Wearable Technologies For Joint Health Assessment

Abstract

Multi-modal sensing relating to joint acoustic emission and joint bioimpedance. Custom-design analog electronics and electrodes provide high resolution sensing of bioimpedance, microphones and their front-end electronics for capturing sound signals from the joints, rate sensors for identifying joint motions (linear and rotational), and a processor unit for interpretation of the signals. These components are packed into a wearable form factor, which also encapsulates the hardware required to minimize the negative effects of motion artifacts on the signals.

Claims

1. A system for assessing joint health comprising: a first sensing assembly for sensing characteristics related to joint physiology at a first time and a second time; a second sensing assembly for sensing characteristics related to joint structure at a first time and a second time; and a health assessor that provides an assessment of joint health through interpretation of characteristics from the first and the second sensing assemblies at the first time, and at the second time; wherein in between the first time and the second time, the joint is perturbed, such that characteristics from the first and the second sensing assemblies at the first time represent the joint pre-perturbation and the characteristics from the first and the second sensing assemblies at the second time represent the joint post-perturbation.

2. The system of claim 1, wherein the perturbation comprises an activity performed by the joint.

3. The system of claim 1, wherein the perturbation comprises changes to the environment of the joint enforced externally.

4. The system of claim 1, wherein the perturbation comprises changes to the environment of the joint enforced internally.

5. The system of claim 1, wherein the perturbation comprises changes to joint conditions enforced externally.

6. The system of claim 1, wherein the perturbation comprises changes to joint conditions enforced internally.

7. The system of claim 1, wherein the first sensing assembly comprises a first wearable sensor for placement proximate the joint; and wherein the second sensing assembly comprises a second wearable sensor for placement proximate the joint.

8. A system for predicting joint health comprising: a first wearable sensor for placement proximate the joint for sensing characteristics related to joint physiology at a first time and a second time; a second wearable sensor for placement proximate the joint for sensing characteristics related to joint structure at a first time and a second time; a health assessor that provides a prediction of joint health through interpretation of characteristics from the first and the second sensing assemblies at the first time, and at the second time; and an output assembly capable of providing a prediction of joint health to a user of the system; wherein in between the first time and the second time, the joint is perturbed, such that characteristics from the first and the second sensing assemblies at the first time represent the joint pre-perturbation and the characteristics from the first and the second sensing assemblies at the second time represent the joint post-perturbation.

9. The system of claim 8, wherein the health assessor uses clustering analysis to compare the pre- and post-perturbation characteristics from the first and the second sensing assemblies.

10. The system of claim 8, wherein the health assessor uses graph mining techniques to compare the pre- and post-perturbation characteristics from the first and the second sensing assemblies.

11. The system of claim 8, wherein the first wearable sensor comprises a wearable acoustic sensor to measure acoustic emissions from the joint.

12. The system of claim 8, wherein the first wearable sensor comprises a wearable sensor to measure at least one non-acoustic characteristic of the joint.

13. The system of claim 8, wherein the second wearable sensor comprises a wearable sensor to measure bioimpedance of the joint.

14. The system of claim 13, wherein the second sensing assembly comprises a plurality of second wearable sensors to measure bioimpedance of the joint in the form of surface electrodes configured to measure proximal electrical bioimpedance of tissue and blood in proximity of the joint.

15. A system for predicting joint health comprising: a first sensing assembly comprising a first wearable sensor for placement proximate the joint for sensing characteristics related to joint physiology at a first time and a second time; a second sensing assembly comprising a second wearable sensor for placement proximate the joint for sensing characteristics related to joint structure at a first time and a second time; a health assessor that provides a prediction of joint health through interpretation of characteristics from the first and the second sensing assemblies at the first time, and at the second time; and an output assembly capable of providing a prediction of joint health to a user of the system; wherein in between the first time and the second time, the joint is perturbed, such that characteristics from the first and the second sensing assemblies at the first time represent the joint pre-perturbation and the characteristics from the first and the second sensing assemblies at the second time represent the joint post-perturbation; and wherein the perturbation is selected from the group consisting of exercise, changing skin temperature, changing joint stiffness conditions, and subjecting the joint to an electromagnetic wave.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0101] FIG. 1 is a block diagram of knee joint acoustic emissions sensing and interpretation for quantifying joint health during rehabilitation according to an exemplary embodiment of the present invention.

[0102] FIGS. 2(a)-(b) illustrate sensor placement and measurement block diagrams according to an exemplary embodiment of the present invention.

[0103] FIGS. 3(a)-(d) illustrate joint sound processing of recordings taken with an electret microphone positioned at the lateral side of the patella (a-c) and results (d) according to an exemplary embodiment of the present invention.

[0104] FIGS. 4(a)-(b) illustrate joint sounds simultaneously sensed by electret, MEMS, and piezoelectric film microphones during three repetitions of (a) flexion/extension and (b) sit-to-stand exercises according to an exemplary embodiment of the present invention.

[0105] FIG. 5 illustrates joint sounds measured on the skin and 5 cm off the skin during flexion/extension exercises according to an exemplary embodiment of the present invention.

[0106] FIG. 6(a) is a block diagram of a bioimpedance measurement system according to an exemplary embodiment of the present invention. FIG. 6(b) is an algorithm for identifying the time intervals when the user is in the correct position to acquire measurements (sitting, legs extended and supported) according to an exemplary embodiment of the present invention.

[0107] FIG. 7(a) show static and dynamic resistance signals (r(t) and r(t)) measured from a subject during various activities according to an exemplary embodiment of the present invention. FIG. 7(b) is a segment of the dynamic resistance signal, zoomed in at * (see FIG. 7(a)). FIG. 7(c) is an ensemble averaged dynamic resistance signals while the subject assumes different positions. FIG. 7(d) illustrates the means and the standard deviations of four features extracted from r(t) shown in bar graphs for the accepted and rejected measurement time intervals.

[0108] FIG. 8 is a block diagram of the bioimpedance measurement system for local joint health assessment according to an exemplary embodiment of the present invention.

[0109] FIG. 9 is another block diagram of the bioimpedance measurement system for local joint health assessment according to an exemplary embodiment of the present invention.

[0110] FIG. 10 is a physiology-driven algorithm design for extracting cardiovascular parameters automatically from the dynamic component (i.e. IPG) of the bioimpedance signals according to an exemplary embodiment of the present invention.

[0111] FIGS. 11(a)-(d) illustrate circuit verification. (a) Photo of the printed circuit board (PCB) fabricated for the analog front-end. (b) Calibration of the resistance and reactance measurements demonstrated high linearity and consistency (points are shown for multiple measurements taken on different days). (c) Noise spectral density plots for dynamic impedance measurements (i.e. IPG). The total noise integrated from 0.8-20 Hz was found to be 18 .sub.rms for the resistive and 55 .sub.rms for the reactive dynamic components, lower than any other previously reported design in the existing literature. (d) Demonstration of improvement of calibration using the innovative microcontroller-enabled automatic calibration methods according to an exemplary embodiment of the present invention.

[0112] FIGS. 12(a)-(e) illustrate human subject measurements according to an exemplary embodiment of the present invention. (a) Electrode placement for knee IPG acquisition. (b) Static impedance measurements. (c) Example waveforms from the intermediate stages of the preprocessing and the feature extraction. (d) Comparison of extracted features using ECG-based vs. IPG-based ensemble averages. (e) Evaluation of the system in response to changes in physiology (vasoconstriction after 2 minutes cold pressor test).

[0113] FIG. 13 is a graph of the magnitude of difference in resistance versus reactance between the left and right leg for control subjects (circle markers) versus following acute knee injury (square markers) according to an exemplary embodiment of the present invention. This small feasibility study already demonstrated a statistically significant difference between the populations, as evident by the clear separation between the groups, and in accordance with prior studies in the existing literature with bench-top bioimpedance measurement circuits.

[0114] FIG. 14 illustrates the concept of acquiring two sets of recordings with the system, the first taken as a baseline recording, and the second taken following a perturbation.

[0115] FIG. 15 is an illustration of the effects of vertical loading forces on the acoustic emissions resulting from frictional rubbing of internal surfaces in the knee.

[0116] FIG. 16 is an illustration of the measurement setup, with the subject performing vertical leg press at varying weights up to the body weight, and the measurements of acoustical emission.

[0117] FIGS. 17-18 illustrate sensor placement and overview of the method of how the signals acquired are analyzed. In FIG. 17, four contact microphones are placed on the medial and lateral sides of the patella and superficially to the medial and lateral meniscus. In FIG. 18, the signal analysis workflow for knee joint sounds. The signals from the dominant knee of the subjects are filtered and standardized (to zero mean and unity variance) and windowed (frame length of 200 ms with 90% overlap). The features are extracted for all four mics and vertically concatenated where columns represent the features and rows represent all the windowed segments. A k-Nearest Neighbor graph (kNN graph) is constructed from the matrix formed using data from the dominant knee and calculates the graph community factor (GCF) using the graph community detection algorithm.

[0118] FIGS. 19-20 show an alternative exercise that can measure joint sounds with different loads. FIG. 19Alter G device allows the user to reduce their body weight by a designated percentage down to 20 percent. FIG. 20 is an illustration of the activity and measurements of acoustical emissions from the joint with contact microphones with respect to the angle of the dominant knee.

[0119] FIG. 21 are graphs created based on the sound features for all windows of the recording and calculating the GCF score. Example graph from one subject are shown with the associated loading condition and GCF.

[0120] FIG. 22 is a box plot showing GCF increases with loading for subject (n=11), indicating more heterogeneity for all acoustical signatures. The asterisk (*) represents the p-value less than 0.01 which is calculated using a non-parametric paired Kolmogorov-Smirnov test.

[0121] FIG. 23 is a box plot showing GCF for different microphones (locations on the knee). The data matrix consists of all loading conditions per microphone for each subject. The asterisk (*) indicates the p-value less than 0.01 which is calculated using a non-parametric paired Kolmogorov-Smirnov test.

[0122] FIG. 24 are graphs related to the set-up of FIGS. 19-20, being graphs created based on the sound features for all windows of the recording and calculating the GCF score. Example graph from one subject is shown with different percentage of the body weight.

[0123] FIG. 25 is an illustration of concept for how physiological biomarker driven recovery can ensure that a player remains in the optimal zone (or at least outside the zone where there is a high risk of overuse injury) by quantifying when the elbow and shoulder joint health is sufficiently recovered.

[0124] FIGS. 26(a)-(d) illustrate (a) an overview of sensing technologies and hardware, (b) example movements used to generate acoustical emissions from the joints, (c) joint angle (top) and acoustical emission (bottom) signals measured in preliminary work, and (d) machine learning based approach to quantifying recovery/fatigue of a joint.

DETAIL DESCRIPTION OF THE INVENTION

[0125] To facilitate an understanding of the principles and features of the various embodiments of the invention, various illustrative embodiments are explained below. Although exemplary embodiments of the invention are explained in detail, it is to be understood that other embodiments are contemplated. Accordingly, it is not intended that the invention is limited in its scope to the details of construction and arrangement of components set forth in the following description or illustrated in the drawings. The invention is capable of other embodiments and of being practiced or carried out in various ways. Also, in describing the exemplary embodiments, specific terminology will be resorted to for the sake of clarity.

[0126] It must also be noted that, as used in the specification and the appended claims, the singular forms a, an and the include plural references unless the context clearly dictates otherwise. For example, reference to a component is intended also to include composition of a plurality of components. References to a composition containing a constituent is intended to include other constituents in addition to the one named.

[0127] Also, in describing the exemplary embodiments, terminology will be resorted to for the sake of clarity. It is intended that each term contemplates its broadest meaning as understood by those skilled in the art and includes all technical equivalents which operate in a similar manner to accomplish a similar purpose.

[0128] Ranges may be expressed herein as from about or approximately or substantially one particular value and/or to about or approximately or substantially another particular value. When such a range is expressed, other exemplary embodiments include from the one particular value and/or to the other particular value.

[0129] Similarly, as used herein, substantially free of something, or substantially pure, and like characterizations, can include both being at least substantially free of something, or at least substantially pure, and being completely free of something, or completely pure.

[0130] By comprising or containing or including is meant that at least the named compound, element, particle, or method step is present in the composition or article or method, but does not exclude the presence of other compounds, materials, particles, method steps, even if the other such compounds, material, particles, method steps have the same function as what is named.

[0131] It is also to be understood that the mention of one or more method steps does not preclude the presence of additional method steps or intervening method steps between those steps expressly identified. Similarly, it is also to be understood that the mention of one or more components in a composition does not preclude the presence of additional components than those expressly identified.

[0132] The materials described as making up the various elements of the invention are intended to be illustrative and not restrictive. Many suitable materials that would perform the same or a similar function as the materials described herein are intended to be embraced within the scope of the invention. Such other materials not described herein can include, but are not limited to, for example, materials that are developed after the time of the development of the invention.

Methods for Sensing Acoustical Emissions from the Knee for Wearable Joint Health Assessment

[0133] In an exemplary embodiment of the present invention, wherein the multi-modal sensing relates to one type of joint characteristic, like joint acoustic emission, the invention can comprise systems and methods of wearable joint rehabilitation assessment following musculoskeletal injury using miniature sensors readily integrated into a wearable device enabling, for the first time, wearable joint acoustics sensing (FIG. 1).

[0134] Microphone Selection

[0135] When investigating types of sensors to use in the present invention, the following were considered: (i) their ability to sense acoustic emissions, and (ii) their practicality for integration within a wearable system. Analysis of how joint sounds propagate through the tissue and transmit to the air suggest contact microphones are the most appropriate sensor for acquiring joint sounds, and a review of conventional systems illustrated that most research employed contact microphones successfully in clinical/lab applications. A contact microphone should theoretically acquire the highest quality acoustic signal since it senses the original, non-attenuated signal and is not sensitive to background noise.

[0136] However, during motion and unsupervised at-home activity, loss of the sensor-to-skin interface is likely and of significant concern, for any compromise to the interface will be detrimental to the signal. In the extreme case that the sensor loses contact with the skin, the system will be unable to record joint sounds completely.

[0137] To improve robustness, air microphones provide complementary sensing capabilities. The signal obtained by air microphones is inherently different from contact microphones. Air microphones will only detect the airborne sounds: attenuated, higher frequency signals. Additionally, while not limited by the sensor-to-skin interface like contact microphones, air microphones are much more susceptible to background noise. For these reasons, the present invention employed both sensing modalities-contact and air microphones-to more robustly capture the acoustic emissions from the joint in a wearable device.

[0138] For the contact microphone, a piezoelectric film (SDT, Measurement Specialties, Hampton, Va., US) was selected because its form factor seemingly lends itself to a wrap and other devices conventionally worn on the knee. Furthermore, piezoelectric films have wider bandwidths compared to miniature, low cost accelerometers, allowing for sensing of high frequency audio signals.

[0139] Two types of air microphones were selected to supplement the piezoelectric film in acquiring acoustic emissions from the knee joint. The first was a commercially-available electret microphone (Sanken Microphone Co., Ltd., Japan). The second was a MEMS microphone, specifically the MP33AB01 (STMicroelectronics, Geneva, Switzerland), which was mounted on a custom PCB.

[0140] Electret and MEMS microphones sense sounds in a similar manner; however, the commercial electret microphone is much more expensive (100) compared to the MEMS microphone. The MEMS's low-cost and sensing capabilities provide a more realistic solution for implementation in a wearable device; however, both the electret and MEMS microphones were used during experiments with the electret microphone acting as the industry standard in terms of the quality of the sound acquired. Recordings from the air microphones were of primary focus because presently they provide higher quality recordings.

[0141] Methods for Microphone Comparison

[0142] The similarity of the MEMS and electret microphones in detecting knee joint acoustic emissions was quantified by computing the information radius between the normalized histograms of these signals, which were acquired by both sensors at the same time placed in the same location on the lateral side of the patella. To construct the aforementioned histograms, the signals acquired from the microphones were first normalized such that their amplitudes were limited to the range [0, 1]. The histogram was formed from this normalized signal using 1000 bins.

[0143] Next, the quality of each sensor was evaluated by computing the signal-to-noise-and-interference ratio (SNIR). The SNIR for each microphone was calculated by finding the ratio of the peak power of a click (i.e., acoustic emission) emitted by the knee joint to the peak power of interface noise in the vicinity of the click. For this calculation, acoustic emissions from the microphones positioned at the medial side of the patella for the air microphones and distal side of the patella for the contact microphone were used.

[0144] Lastly, a proof-of-concept experiment was conducted to compare signals measured on and off of the skin. A subject performed three cycles of seated, unloaded knee flexion/extension with two electret microphones positioned at the lateral side of the patella, one on the skin and one located 5 cm off the skin. The resulting signals were then compared.

[0145] Interfacing Circuits

[0146] The analog front-end for the MEMS microphones comprised a non-inverting amplifier stage with 33 dB gain, which was selected such that the signals do not saturate but are amplified to utilize the full dynamic range of the subsequent analog-to-digital converter, and a high-pass 15 Hz cutoff frequency. This stage was followed by a second-order low-pass filter with a cutoff frequency of 21 kHz. A bandwidth of 15 Hz-21 kHz was chosen, as knee joint sounds can range between these frequencies.

[0147] The analog front-end for the piezoelectric film microphones comprised an amplification stage of gain 45 dB and 100 Hz high-pass cut off. This stage was followed by a fourth-order low pass filter with a 10 kHz cut-off frequency. A 100 Hz high-pass cut off was chosen to attenuate the interface and motion artifact noise.

[0148] Human Subject Study and Measurement Protocol

[0149] Thirteen male subjects without history of knee injuries participated in the study and gave written informed consent approved by the Georgia Institute of Technology Institutional Review Board (IRB) and the Army Human Research Protection Office (AHRPO). The subject population was reasonably homogenous in terms of physical activity level (collegiate athletes) and ranged in age (19-21 years), weight (84.1-135.3 kg), and height (174-195 cm). With this approach, the plan was to assess the variability in the measurements separately from variability due to age or knee joint health.

[0150] Following preliminary measures of body composition, height, and weight, an electret and MEMS microphone were both positioned at the lateral and medial sides of the subject's patella targeting the patellofemoral joint while two piezoelectric film sensors were placed on the skin just proximal and distal to the patella. Each sensor was attached using Kinesio Tex tape. In addition to the tape, a thin piece of silicone (5 mm thick) was placed over the piezoelectric film to reduce the interface noise of the tape rubbing against the film. Lastly, two wireless inertial measurement units (IMUS) (MTW-38A70G20, Xsens, Enschede, The Netherlands), which contained three-axis accelerometer, gyroscope, and magnetometer as well as built-in sensor fusion outputs, were positioned on the lateral sides of the thigh and shank.

[0151] FIG. 2 illustrates sensor placement and measurement block diagram. In FIG. 2(a), eight sensors were used during human subject testing. Two IMUS were placed laterally on the thigh and shank. Piezoelectric film sensors were placed directly proximal and distal of the patella. The air microphones (MEMS and electret) were attached on the lateral and medial sides of the patella. FIG. 2(b) is a block diagram of the data collection hardware used during human subject studies.

[0152] While wearing these sensors, each subject completed two exercises: (i) seated, unloaded knee flexion/extension and (ii) sit-to-stand. For each exercise, the subject repeated the motion five times while the microphone and IMU outputs were recorded in a quiet room (FIG. 2(b)). The signals from the piezoelectric and MEMS microphones were passed through custom circuits and then collected at 50 kHz (16 bits/sample) using Biopac data acquisition hardware (Biopac Systems Inc, Goleta, Calif., US) while the signals from the electret microphones were sampled at 44.1 kHz (16 bits/sample) using a Zoom H6 recorder (Zoom Corp., Tokyo, Japan).

[0153] The IMU signals were acquired at 50 Hz (16 bits/sample) using their device-specific software suite (MT Manager, Xsens, Enschede, The Netherlands) synched with the Biopac system. Apart from the electret microphone signals, which were stored on an SD card (SanDisk, Milpitas, Calif., US) via the Zoom recorder, all signals were recorded on a laptop. The data were then processed using MATLAB (The Mathworks, Natick, Mass.).

[0154] Joint Sound Processing

[0155] The signal processing comprises (i) calculation of knee joint angle and contextualization of the joint sounds with joint angle, (ii) identification of significant high frequency acoustic emissions or clicks, and (iii) statistical analysis to quantify the consistency of occurrence of the main clicks with respect to joint angle.

[0156] First, the knee joint angle was calculated using the methods that leverage the sensor fusion outputs of 3-axis accelerometer, gyroscope, and magnetometer provided by Xsens, namely the rotation matrix (i.e., Direction Cosine Matrix), and the kinematic constraints of a hinge joint to provide angle data. This method allowed for arbitrary sensor placement and orientation on each segment of the joint (i.e., thigh and shank), eliminating the need for precise calibration techniques and measures. However, this method is potentially susceptible to error, due to deviations from a true hinge joint as a result of skin and motion artifacts. Nevertheless, since the cycles of repetitive motions were analyzed against one another, this error was common to each cycle and thus did not present in the results.

[0157] Finally, the signal was normalized between 0 and 90 such that subjects could be compared against one another with respect to location within each subject's range of motion.

[0158] FIGS. 3(a)-(d) illustrate joint sound processing of recordings taken with an electret microphone positioned at the lateral side of the patella (a-c) and results (d). FIG. 3(a) is an example 3000 sample (60 ms) joint sound recording window showing three distinct high-amplitude, short-duration, acoustic emissions. The original signal contains ambient noise, which presents as broadband signals up to 7 kHz, and interface noise, which appears as baseline movement. These components are clearly visualized in the spectrogram of the original signal. To remove the majority of the noise, the signal is bandpass filtered at 7 kHz-16 kHz, resulting in the filtered signal x[n].

[0159] The envelope of this signal is found, yielding A [n]. Using a thresholding technique based on the moving average, the significant peaks of A [n] are found, roughly corresponding to the clicks of the original signal. These are later refined to match the true locations of the clicks found in the original signal (i.e., such that the locations correspond to where the clicks achieve their maximum amplitudes, positive or negative, in the original signal). FIG. 3(b) illustrates a final result of the click detection algorithm, which displays the identified clicks for three cycles of flexion/extension. FIG. 3(c) illustrates three extension cycles with artificial offsets. These qualitatively show that the main acoustic event of each cycle occurs at similar angular locations. FIG. 3(d) illustrates the final results of click location consistency for five repetitions of flexion/extension for 13 subjects on left and right legs. Across subjects, the standard deviation for click location is small, supporting observations of consistent angular location cycle-to-cycle. Additionally, the mean locations of these clicks are consistent between left and right legs for most of the subjects.

[0160] As discussed and in more detail, once knee joint angle was calculated, the phases (flexion or extension) of each cycle were determined (FIG. 3(b)). First, the inflection points of the signal were found. The beginning and ending of each phase were identified by finding pairs of inflection points. These points were then used to segment the microphone signals, contextualizing the data by type of angular motion.

[0161] Next, significant acoustic emissions were identified. The most distinct audio signals that were detected by the air microphones were the high amplitude, short duration clicks (FIG. 3(a)). By observing the signal's frequency content (i.e., short-time Fourier transform, STFT), these clicks were broadband with frequencies as high as 20 kHz. This unfiltered signal contained two main sources of noise; ambient noise ranged in frequency up to 7 kHz while interface noise appeared as baseline movement with components up to 1.5 kHz. The first step of this identification stage was to preprocess the signal such that the clicks became more prominent and any interface and/or ambient noise were mostly cancelled. To this extent, the air microphone signals were filtered with a bandpass filter spanning 7 kHz-16 kHz. As seen from FIG. 3(a), the filtered signal (x[n]) lacks the original baseline movement, and the clicks are more distinct from other artifacts in the signal.

[0162] After this preprocessing step was complete, a modified envelope detection algorithm was implemented. A 1024-bin spectrogram of the signal (X[n, m]) was calculated with a window size of 100 samples (i.e., 2 ms) and 90% overlap. The amplitude of the signal was calculated by summing the logarithmic amplitude of the spectrogram across the frequency bins as follows:

[00001]$\begin{array}{cc}A\left[n\right]=\underset{j}{.Math.}.Math.20.Math..Math.\log .Math..Math.X\left[n,j\right].Math.& \left(1\right)\end{array}$

[0163] A moving average and standard deviation ([n] and [n]) of A[n] using a window size of 1000 samples were calculated. A[n] was then thresholded such that:

[00002]$\begin{array}{cc}T\left[n\right]=\{\begin{array}{cc}A\left[n\right],& A\left[n\right]>\left[n\right]+.Math.\left[n\right]\\ 0,& \mathrm{otherwise}\end{array}& \left(2\right)\end{array}$

[0164] where T[n] is the thresholded amplitude signal and is a constant control coefficient, which was selected as 3.3 by inspection.

[0165] Next, the peaks of T[n] were detected by standard peak detection techniques. The peaks that resulted from the same click (i.e., resonances of the initial click, which are specified as peaks within 150 samples of each other) were eliminated, resulting in the raw click locations vector p.sub.r=[p.sub.r1,p.sub.r2, . . . ,p.sub.rL]. The raw click locations p.sub.r were refined such that each click location corresponded to the point on the original filtered signal where the click achieved its maximum amplitude, positive or negative. The refined click locations matrix p=[p.sub.1,p.sub.2, . . . ,p.sub.L] gave the final detected click locations. An example of these detected clicks is shown in FIG. 3(b).

[0166] Once the clicks were identified, the consistency of these acoustic emissions was analyzed. FIG. 3(c) provides a visualization of consistent acoustic emission during repetitive motion. For each cycle of a particular exercise (i.e., flexion or extension), the three clicks with the largest amplitudes and their corresponding angular locations were determined. Each combination of the clicks across cycles (i.e., selection of one of the three clicks from each cycle) was found. The combination with the smallest standard deviation for angular location yielded the most consistently occurring major acoustic event. The mean and standard deviation of these locations were calculated.

[0167] Given these mean locations, three methods were used to analyze the data. For the first two methods, test-retest reliability was estimated using the ICC. The data was organized into motions and repetitions. There were 52 motions, one for each human subject and exercise combination (e.g., subject 1's extension data represented one motion). The repetitions comprises the five click locations (one per cycle) from the selected combination. This dataset will be referred to as the test-retest dataset.

[0168] Given this dataset, two ICC values were calculated using one-way random single (i.e., ICC(1, 1)) and average measure (i.e., ICC(1, k)) models to show the reliability of a single cycle's measure and mean of the fives cycles' measures. Additionally, the 95% confidence intervals (CI) for these two ICC values were determined. The last method for analyzing the data was a paired t-test, which was used to assess whether there were significant differences between the mean click locations for left and right legs.

Results and Discussion

[0169] Microphone Comparison

[0170] In evaluating microphone selection, many different parameters were considered. First, the similarity of the signals measured by the electret and MEMS microphones were compared. The quality of these microphones was determined by evaluating the quality of their sensing capabilities in terms of SNIR. Moreover, when investigating interface issues for the air microphones, the effect that the sensor-to-knee distance had on the signal acquired was examined. Finally, the quality of the contact microphone was researched.

[0171] FIG. 4. illustrates joint sounds simultaneously sensed by electret, MEMS, and piezoelectric film microphones during three repetitions of FIG. 4(a) flexion/extension and FIG. 4(b) sit-to-stand exercises. For both parts FIGS. 4(a)-(b), the top plot displays the joint angle ([n]). The middle and bottom graphs show the time and frequency domain signals from the various microphones. The acoustic signatures of the electret and MEMS microphones exhibit similar characteristics.

[0172] As shown in FIGS. 4(a)-(b), the electret and MEMS microphones, measuring frequencies as high as 20 kHz, performed similarly in detecting joint sounds, which were acquired from a subject performing flexion/extension and sit-to-stand exercises respectively. This was confirmed by computing the information radius between the normalized histograms of signals captured by these two microphones, which yielded a value of 0.0025. This value shows a high similarity between these two types of microphones since the information radius ranges from 0 for identical distributions to 2 for maximally different distributions. This shows that the more cost-effective MEMS microphones are a viable substitute for the more expensive electret microphones. This is an important result when designing deployable systems.

[0173] As predicted, the signal recorded by the air microphones included noise and interface components in addition to the desired joint sounds; both ambient background interference and interface noise caused by the rubbing of athletic tape, which was used to hold the sensors in place, were sensed by the microphones. The SNIR was 11.7 dB for the electret microphone and 12.4 dB for the MEMS microphone. To minimize issues with noise during initial experiments, measurements were taken in a quiet room. Further investigation can be addressed for implementation of a deployable, wearable system, especially given the fact that many background noises, such as speech and sounds due to ambulatory motion, will reside in-band with the joint sounds.

[0174] FIG. 5 illustrates joint sounds measured on the skin and 5 cm off the skin during flexion/extension exercises. Though the off-skin microphone captured a signal with decreased amplitude, the on-and off-skin measurements showed significant similarities in their acoustic signatures. The main acoustic event of each signal occurred at similar locations.

[0175] Experiments showed that the air microphones did not need to be directly located at the skin surface to detect airborne joint sounds. As shown in FIG. 5, the sounds obtained from an electret microphone placed on the skin and one located 5 cm off the skin captured similar acoustic signals in both morphology and timing (4.5 smaller amplitude). This is an important observation because it suggests that the air microphones can record joint sounds in a wearable device where direct contact with the skin may not be constant. However, it will be important to consider this distance when analyzing the captured signals, particularly when the analysis depends on the amplitude of the signal.

[0176] In this sense, maintaining a fixed distance between the microphone and skin, especially for use in longitudinal analysis, may be required. Furthermore, placing the microphone off of the skin introduces increased potential for noise; the microphone may have a greater opportunity to strike or rub against the skin. Additionally, changing the distance between the microphone and skin will change the microphone's sensitivity in sensing these sounds.

[0177] The piezoelectric film measured signals up to approximately 3 kHz as seen from the spectrograms of the signals acquired shown in FIG. 4. While the piezoelectric film had the advantage of not detecting background noise, it acquired significantly more interface noise8.4 dB SNIRdue to the sensor rubbing on the skin and the athletic tape rubbing on both the skin and the sensor. This interface noise had frequency components up to 1.5 kHz and was thus in-band.

[0178] During early pilot data collections, the piezoelectric film was attached to the skin using only Kinesio Tex tape. However, this method proved to be very susceptible to interface noise. As the knee extends and flexes, the tape, though stretchable, deformed the film which obscured the low frequency and low amplitude signatures. Furthermore, though acceptable for collecting pilot data, tape proves to be undesirable for long-term monitoring. To mitigate this issue, a piece of silicone was placed above the piezoelectric film. Because silicone has similar compliant mechanical properties to skin and subcutaneous tissue, the joint sounds received did not experience dampening, and the silicone surface provided a suitable surface to stick the tape. Though this method did not completely eliminate interface noise-the sensor still experienced some movement along the skin-it did help to reduce the recorded noise.

[0179] Accordingly, while using piezoelectric film or other contact microphones is desired to capture the vibration signal, which represents the majority of the acoustical energy generated, implementation presents practical issues. The piezoelectric film was significantly affected by interface noise. A smaller portion of the signal bandwidth was corrupted by interface noise for the air microphones compared to contact microphones. Furthermore, contact microphones did not pick up higher frequency vibrations as distinctly as air microphones. For these reasons, the use of air microphones for wearable joint sound measurements might be preferable.

[0180] Joint Sound Consistency

[0181] FIG. 3(d) summarizes the results for mean angular click location for the left and right legs of 13 seated subjects performing five repetitions of knee flexion/extension. Two important findings resulted from this data: (i) significant acoustic events are repeatable during single trial measures, and (ii) left and right legs produce similar sounds.

[0182] First, two ICC values were found for the test-retest dataset. An ICC(1,1) value of 0.94 with a 95% CI of 0.92-0.97 and an ICC(1, k) value of 0.99 with a 95% CI of 0. 98-0.99 were calculated. Since the ICC values were greater than 0.7, these values showed that the main acoustic emission per cycle of activity were consistent within a single trial of monitoring for both single and average measure reliability. Given that audible joint sounds have not been extensively explored, this was an important finding, demonstrating that airborne signals emit a stable pattern with repeated movement in a healthy hinge joint.

[0183] Second, the difference between legs for each exercise suggested that a healthy subject's knees produce similar joint sounds. The difference between left and right legs were not significant at the p<0.05 level. While as a group, there were no significant differences between the left and right legs, some subjects could be grouped as having relatively no difference between right limb and left limb click locations whereas others had notable differences between right and left suggesting the potential for defining clinically relevant signature traits. Such variations in click location could represent useful knee joint health biomarkers.

[0184] Though these results are promising, there are some limitations to the tested system and analysis. First, with regard to the IMUs, sensor positioning, drift, and motion artifacts can all contribute to flexion angle calculations that differ from the true joint angle. Techniques discussed herein, and others will need to be employed to minimize these errors, especially when considering their application in a system which measures longitudinal data. For example, some errors could be minimized by ensuring more rigid sensor positioning and leveraging the joint's kinematic constraints directly into the calculation of joint angle to reduce the effect of drift.

[0185] Second, the effect of lubrication (e.g., diminished boundary lubrication after an injury) and differing structural components (e.g., damaged ligaments, etc.) on acoustic emissions has not yet been sufficiently studied. These variables may introduce error when calculating click location consistency for repeated cycles and measuring differences between legs. In this sense, these isolated, one-time measurements may not prove to be as useful as compared to longitudinal analysis for the same subject over time.

[0186] These quantitative findings in terms of measurement consistency form the foundation for understanding the significance of changes in joint sound signatures associated with injury, as well as changes in such signatures during rehabilitation. Moreover, while longitudinal studies will be important towards understanding injury recovery, this work and its focus on robust implementation in a wearable platform also presents opportunities for exploring day-to-day and within-day changes of joint acoustics.

CONCLUSIONS

[0187] In an exemplary embodiment of the present invention, wherein the multi-modal sensing relates to knee joint acoustic emission during loaded and unloaded activities, it is demonstrated, quantitatively, that major acoustic events occur at consistent joint angles during repetitive motions for healthy subjects. Furthermore, these locations are similar between left and right legs for most subjects. Whether asymmetry between right and left knee acoustic emissions is related to risk factors for injury or other training-related variables remains to be clarified. Importantly, these findings showed that joint sound measurements from air microphones are repeatable with sensing technology that can be implemented in a relatively inexpensive, wearable form factor. While extensive analysis of the piezoelectric film was not conducted, its use in a wearable device holds promise based on preliminary findings showing that packaging techniques have a large influence on the signal recorded.

[0188] The present invention further includes mitigating background and interface noise for both the air and contact microphones. In particular, a focus on the packaging of these sensors into a wearable wrap or sleeve enabling high quality signal measurements during at-home, long-term monitoring is investigated. Additionally, existing algorithms are refined, and new processing techniques developed to detect clinically-relevant acoustic signatures.

[0189] Given that therapists and clinicians look at sounds, swelling, structural stability, and range of motion, the present invention investigates methods for quantifying these joint health biomarkers unobtrusively and accurately; namely, it determines which acoustic signatures encapsulate these biomarkers. Furthermore, exploration of these biomarkers as they relate to specific diseases and injuries (e.g., osteoarthritis, anterior cruciate ligament tear, meniscal tear, etc.) are considered. Finally, longitudinal studies on injured subjects allow for the determination and validation of specifics acoustic emission features (e.g., consistent angular location) that provide valuable joint health information during rehabilitation following an acute injury.

A Robust System for Longitudinal Knee Joint Edema and Blood Flow Assessment Based on Vector Bioimpedance Measurements

[0190] In another exemplary embodiment of the present invention, joint characteristics including edema and blood flow parameters are examined and systems and methods of wearable joint rehabilitation assessment following musculoskeletal injury using bioimpedance technologies readily integrated into a wearable device.

[0191] In this embodiment of the present invention, systems and methods are disclosed that address technological gap in the area of wearable bioimpedance measurement systems for local joint physiology assessment.

[0192] Position Identification Algorithm

[0193] Bioimpedance measurements are too greatly impacted by motion artifacts, subject position, electromagnetic interference and voltage fluctuations of the skin electrode interference. Therefore, for consistency, measurements should be taken when subject is still in a given position and in the absence of electromagnetic interference and skin electrode interface related fluctuations (such as an electrode losing contact with the skin). These conditions can be met under user guidance however this kind of guidance is not feasible in a wearable device setting.

[0194] As discussed previously, IMUs can be used to decide if the user is still and in an acceptable position for measurements to be taken. However, IMUs will not be effective in detecting electromagnetic interference or skin-electrode interface related fluctuations. These kinds of effects can however be detected using the dynamic resistance signal which is greatly influenced by them, as well as the user position and motion artifacts.

[0195] IMUs along with the dynamic resistance (impedance plethysmography) signals can be used together to decide if the user is in an acceptable position for bioimpedance measurements to be taken or not. The IMUs would be used to provide information about the subject's limb position (such as knee angle) and the activity that the subject is performing (such as walking, running, sitting still etc.). The dynamic resistance signal can be used to assist in detecting the subject's limb position and the activity being performed. It can also be used to deduce if any of the electrodes were misplaced and if any of the electrodes lost contact with the skin. If any changes on the electrode positioning occur, these can be detected through features extracted from the dynamic resistance signal. The bioimpedance measurements also get affected by muscle contractions. The dynamic resistance signal can also be used to deduce if the limb muscles are relaxed or contracted.

[0196] The dynamic resistance signal was used to design an exemplary algorithm that can decide if the user is in an acceptable position for bioimpedance measurements to be taken. IMUs can be used to assist such an algorithm as well. In this exemplary embodiment, only the dynamic resistance signal was used. In this case, the acceptable position was when the subject is motionless, seated with legs fully extended and supported from the bottom. All other positions or activities (sitting legs bent, standing, any kind of motion) were to be marked as rejected measurements. The static impedance measurements were not used in the decision process as these are the measurements to be interpreted and using them in the decision process would create tendency to acquire data within a certain range of values.

[0197] FIG. 6(a) is a block diagram of a bioimpedance measurement system according to an exemplary embodiment of the present invention. E1-E4 represent the electrodes that interface to the body. The signals i(t) and q(t) relate to the static (slowly varying on the order of hours to days) component and the signals) i(t) and q(t) relate to the dynamic component of the knee impedance. A(t) is used to monitor the amplitude of the current (i.sub.body(t) passing through the knee joint.

[0198] The position identification algorithm is summarized in FIG. 6(b) and used for identifying the time intervals when the user is in the correct position to acquire measurements (sitting, legs extended and supported). The acceptable time intervals are identified by extracting features from the dynamic voltage signal i[n] and using these features to make a decision. The binary decision rule is trained once before-hand. The accepted time intervals are used to obtain the knee resistance (R.sub.measured) and reactance (X.sub.measured) using the static voltage measurements i[n] and q[n]. A[n] is used to cancel the effects of any variations in injected current amplitude.

[0199] The signal i[n] is linearly filtered (0.3 Hz-20 Hz) to get rid of respiratory artifacts and high frequency noise. Then, the filtered signal is windowed with a window size of 10 seconds and a step size of 1 second (decided heuristically) creating M frames of the filtered signal. The signal in each frame is amplitude corrected using A[n] and calibrated to give the dynamic resistance signal within the frame.

[0200] Features extracted from each frame are placed in a feature matrix F=[f.sub.1f.sub.2 . . . f.sub.M].sup.T, where each column corresponds to a feature and each row to a frame. These features are then used to decide whether a frame should be accepted (classified with label 1 or rejected classified as 0), producing a vector of binary labels d=[d.sub.1d.sub.2 . . . d.sub.M].sup.T. Frames with standard deviation of the dynamic resistance signal exceeding 50 m are labeled 0 because it is highly likely they contain motion.

[0201] The static voltage measurements i[n] and q[n] are also windowed using the same scheme, amplitude corrected using A[n] and calibrated producing the resistance and reactance signals for each of the M frames. The mean of the resistance and reactance signals for each frame are taken to give the average resistance and average reactance vectors, r=[r.sub.1 r.sub.2 . . . r.sub.M.sub.].sup.T and x=[x.sub.1x.sub.2 . . . x.sub.M].sup.T respectively. These M resistance and reactance measurements are either accepted or rejected according to the labels in vector d. The accepted resistance and reactance measurements are placed into the accepted resistance and accepted reactance ({tilde over (x)}) vectors respectively. The mean of {tilde over (r)} and {tilde over (x)} are taken to give the final measured resistance (R.sub.measured) and reactance (X.sub.measured).

[0202] The binary decision rule mentioned is trained separately before it is applied to new data (a testing set). For this training, i[n] is recorded while the subject does activities with known labels for a known amount of time: (1) standing (label 0), (2) sitting legs bent (label 0), (3) sitting legs crossed (label 0), (4) sitting legs extended and supported (label 1) and (5) walking (label 0) each for 1 minute.

[0203] Feature extraction is performed on i[n] as shown in FIG. 6(b) to produce a feature vector F.sub.training and the corresponding known labels d.sub.training. Feature subset selection is performed on the features in F.sub.training via classifier subset evaluation using a support vector machine (SVM) classifier with a genetic search algorithm. The selected subset of features and d.sub.training are used to train an SVM (with a cost parameter of 1.0) resulting in the binary decision rule to be used. Note that F.sub.training and d.sub.training do not include those frames where the standard deviation of the dynamic resistance signal exceeds 50 m (frames with motion artifacts). The feature subset selection and SVM training were carried out using Weka while all other tasks were done on MATLAB.

[0204] The features extracted from each frame are described in three categories. The first category of features is generic features that were derived from audio signal feature extraction techniques. The 10-second frame is further partitioned into sub-frames of size 700 ms with 350 ms step size. Temporal and FFT related features (30 features) are extracted from each 700 ms sub-frame, these are the short-term features. Statistics also referred to as mid-term statistics, such as mean, median and standard deviation (8 statistics) are computed for each feature across all the 700 ms sub-frames resulting in a total of 240 features per frame.

[0205] The second set of features is based on the frequency content of the ensemble average of the dynamic resistance waveform across each frame (4 features). The ensemble average of the waveform across the frame is computed using an algorithm. The last sets of features are temporal features extracted from the ensemble average dynamic resistance signal for each frame (23 features).

[0206] FIG. 7(a) show static and dynamic resistance signals (r(t) and r(t)) measured from a subject during various activities. The time intervals marked when the subject is in the correct position for taking measurements (sitting, legs extended and supported). The measurements taken when the subject is in the correct position are to be accepted, all other measurements should be rejected. The position identification algorithm is tested on this data. FIG. 7(b) is a segment of the dynamic resistance signal, zoomed in at * (see FIG. 7(a)). The pulsatile component of the signal due to blood flow is clearly seen. The characteristics of this signal are used for position identification.

[0207] FIG. 7(c) is an ensemble averaged dynamic resistance signals while the subject assumes different positions. Each ensemble average was taken over a time interval of 40 seconds. The difference in the ensemble average waveforms as the subject assumes different positions can be seen clearly. FIG. 7(d) illustrates the means and the standard deviations of four features extracted from r(t) shown in bar graphs for the accepted and rejected measurement time intervals. The first graph shows the bar graph for (log.sub.10) which is related to the SNR of the signal. The second graph is for the peak to peak amplitude of the ensemble average of the signal. The third graph is the 0-3 Hz band power of the ensemble averaged signal. The final graph is for the Euclidean distance between the ensemble average waveform and a template ensemble average waveform acquired during the one-time training phase while the subject assumes the correct position.

[0208] The static and dynamic resistance signals measured, while a subject performs a variety of activities is shown in FIG. 7(a). The time intervals where the subject is in the correct position (seated, legs extended and supported) for the measurement to be acceptable are shown in red, while all others are shown in black: subject standing, seated with legs bent 90, seated leg crossed, walking). This measurement was made over a course of 12 minutes. It can be seen from this figure that both the dynamic and static resistances are greatly impacted by motion artifacts (transition motion labeled as T, walking), which create variations in the orders of one to tens of Ohms.

[0209] A portion of the dynamic resistance waveform while the subject is in the correct position is shown in FIG. 7(b). While the subject is still, this waveform varies periodically over the course of milliseconds, and the variability is in the order of tens of milliohms. The static components of the signal do not have this periodic variability as it has been filtered out.

[0210] The ensemble averaged dynamic resistance waveforms for three 80-second time intervals where the subject was in different positions (standing, seated legs bent 90, correct position) are shown in FIG. 7(c). This figure points out some features that can be effective in classifying the time intervals with the correct position from the others, such as the peak to peak amplitude of the ensemble averaged r(t) waveform. It can also be seen that the correct position's waveform looks smoother (and thus less noisy) than both the standing and the sitting position's waveforms. This suggests that features related to the SNR, frequency domain and shape of the ensemble averaged r(t) waveform are potentially effective features as well.

[0211] FIG. 7(d) shows the means and standard deviations of four features extracted from the 10s time windows where the subject is in the correct position (accepted measurement windows) and incorrect position (rejected measurement windows) for the 12 minute measurement shown in FIG. 7(a). The first feature (shown on the leftmost bar graph) is the minus logarithm of the cross-correlation coefficient between the even and odd sub-averages (log.sub.10) within the 10 second frame. This feature is inversely related to the SNR of the ensemble average waveform. The second and third features (second and third bar graphs from the left) are the peak to peak amplitude and the 0-8 Hz bandpower of the ensemble averaged r(t) waveform within each 10 second frame. The final feature shown (the rightmost bar graph) is the Euclidean distance between the ensemble averaged r(t) waveform, and a template ensemble averaged r(t) waveform taken during the training phase while the subject was in the current position.

[0212] As predicted from FIG. 7(c), these four features show effective class separability. Lilliefors and Kolmogorov-Smirnov tests performed on each class and each of the features showed that the distributions were normal (p<0.01). A two-sample t-test (with unequal and unknown class variances assumed) showed significant class separability in each feature (p<0.01).

[0213] The training data was used to train a binary classification rule using the method described above. The confusion matrix for the classifier evaluated on the testing data where frames with motion (standard deviation of r(t)>50 m) are not included is seen on TABLE I.

TABLE-US-00001 TABLE I Confusion Matrix for the Position Identification Algorithm Evaluated on the Testing Data N = 405 Predicted Reject (0) Predicted Accept (1) Actual Reject (0) 261 5 Actual Accept (1) 60 79

[0214] The baseline misclassification rate (when the most probable class is always chosen, reject (0) all the frames in this case) was 34.3%. The misclassification rate of the trained classifier was 16% and the precision was 94%. Therefore, although a significant number of frames that should be accepted are rejected by the algorithm, most frames accepted are actually acceptable, which is what matters the most for measurement consistency. The significant number of false negatives is not important as the system only aims to take 5 to 10 measurements (with a 10 second duration) within a day, which makes missing some measurements non-critical. Infrequent measurements are sufficient to monitor knee joint health as the static knee impedance has physiology related variations in the course of ours to days.

[0215] Over the course of 12 minutes of testing when only the actual acceptable frames were used for the binary decision rule, the measured resistance (R.sub.actual or mean ({tilde over (r)}) along with std({tilde over (r)})) was 60.90.6 and the measured reactance was 13.50.7. The impedance measurements when all frames without motion were accepted were 60.54.8 and 13.11.5. The higher standard deviation in the impedance measurements is due to inconsistent measurement positions.

[0216] When the binary decision rule trained was used, the impedance measurements were 60.81.5 and 13.541.0 for resistance and reactance respectively. These values are more consistent (have lower standard deviation) and have lower absolute error compared to accepting every frame without motion, due to more consistent subject positioning.

[0217] These results are a demonstration of how an algorithm can be used to automatically decide to take bioimpedance measurements when the subject is in a certain position, eliminating the need for user guidance. An algorithm requiring a one-time training similar to this can be implemented on a smart phone that wirelessly communicates with the bioimpedance hardware, which is a step towards creating a Smart Brace monitoring knee impedance.

[0218] Bioimpedance Systems and Methods

[0219] A system that measures bioimpedance signals and then extracts both musculoskeletal (tissue resistance and reactance), and cardiovascular (heart rate, local blood volume, and flow rate) parameters (FIG. 8) is disclosed.

[0220] FIG. 8 is a block diagram of the bioimpedance measurement system for local joint health assessment according to an exemplary embodiment. An innovative combination of both musculoskeletal and cardiovascular parameters will be obtained from the knee joint longitudinally during recovery from acute musculoskeletal injury with a wearable device based on the high-performance circuit and system described here.

[0221] These physiological parameters are used to quantify both edema and blood flow during post-injury recovery. The invention advances the state-of-the-art for bioimpedance measurement systems by, among other things, (i) incorporating a custom bioimpedance measurement analog front-end for both tissue impedance (static) and local hemodynamics (dynamic) with the highest resolution given the power consumption and size, compared to similar systems, and/or (ii) employing a self-calibration procedure to minimize drift and inaccuracy due to environmental factors such as temperature, and/or (iii) creating customized physiology-driven algorithms for IPG-based heartbeat detection to alleviate the need for a reference biosignal recording (e.g., an electrocardiogram, ECG) for extracting hemodynamic parameters from the knee.

[0222] The aims of quantifying edema and blood-flow in the post-injury period necessitate a small form factor system obtaining accurate bioimpedance measurements from the body in an energy-efficient manner. Towards reaching that goal, a digitally-assisted analog approach was followed: designing a system benefiting from the advantages of both analog (i.e., low power) and digital (i.e., programmability) domains. The present system uses this approach to perform (i) bioimpedance measurements, (ii) calibration, and (iii) preprocessing and feature extraction.

[0223] The first function, performing the bioimpedance measurements from the body, is achieved by a custom, analog front-end designed with discrete components. A low-power TI MSP430 series microcontroller (Texas Instruments, Inc., Dallas, Tex., US) is used with a micro secure digital (SD) card as a data logger to enable processing of the signals later on a computer. The microcontroller is also used to implement the second function of the system, namely performing calibration, which aims to reduce the measurement error due to environmental changes (e.g., temperature). Feature extraction, is performed using MATLAB software (MathWorks, Natick, Mass.) to extract physiologically-relevant information from the calibrated data stored on a microSD.

[0224] Analog Front-End

[0225] Comprising resistive body fluids (e.g., blood, intra-cellular fluid) and capacitive cell-walls, the EBI of a local joint body region can be modeled, to the first order, as a single RC network. It should be noted that the RC network has a static component related to the total fluid-volume and a dynamic component related to the time-dependent fluid-volume changes (e.g., blood flow periodic with the heart beat). The analog front-end was designed to perform a single-frequency bioimpedance analysis to extract both static and dynamic components of the RC network (block diagram of front-end shown in FIG. 9).

[0226] In FIG. 9, E1-E4 represent the four electrodes on the body used for impedance measurement; i(t) and q(t) represent the static (slowly varying on the order of hours to days), and i(t) and q(t) represent the dynamic (rapidly varying on the order of milli-seconds) bioimpedance components; A(t) is the sensed electrical current delivery to the body being monitored in real-time to ensure safety and enable real-time calibration of the measured data on-board the microcontroller (C).

[0227] The circuit excites the body with a sine wave current at f.sub.0=50 kHz, a frequency that enables current to flow through both extracellular and intracellular fluid paths and therefore is widely used in single-frequency bioimpedance analysis systems. The design incorporates a diode-stabilized Wien-Bridge oscillator generating an 800 mV.sub.pp sinusoidal signal at 50 kHz. The voltage output of the oscillator is converted to current by a high output-impedance, high-bandwidth, voltage-controlled current source (VCCS). The VCCS delivers the current to the series combination of the body-load, Z.sub.Body, and a small and pure resistive load of R.sub.sense=100.

[0228] To cancel out the effects of non-ideal skin-electrode interfaces on the measurements, current is injected to the joint through a quadripolar electrode configuration formed by E1-E4 in FIG. 9, where E1 and E4 are current injection and E2 and E3 are voltage measurement electrodes. Voltage measurements across Z.sub.Body and R.sub.sense are made by means of two separate instrumentation amplifier (IA) stages, IA.sub.body and IA.sub.sense.

[0229] The signal at the output of the IA.sub.body, namely v.sub.Body(t) is used to extract voltage measurements corresponding to both resistive and reactive components of the Z.sub.Body through phase-sensitive detection circuitry. A differential high-pass filter (HPF) with a cut off frequency of 7.2 kHz reduces any electromyogram (EMG) bleed-through from the surrounding muscles. The signal at the output of the IA.sub.sense, namely v.sub.sense(t), is used for (i) generating the clocks that drive the phase-sensitive detection circuitry using comparators, and (ii) monitoring the magnitude of the current delivered to the body, namely I.sub.PP, through an envelope detector formed by a diode and RC network, the output of which is the current monitoring signal A(t) in FIG. 9. The clocks for the in-phase and quadrature-phase sensitive detection are generated from the v.sub.sense(t) and a 90 phase shifted version of the v.sub.sense(t) obtained using an all-pass-filter, respectively. Both clocks are used to alternate analog switches between v.sub.Body(.sub.t) and v.sub.Body(t).

[0230] The signal from each phase-sensitive detector is filtered to give the final output signals. For the very slowly-varying static in-phase and quadrature signals, namely i(t) and q(t), the outputs of phase-sensitive detectors are each filtered with a low-pass filter (LPF) having a cutoff frequency of f.sub.3dB=2 Hz. To extract the small-magnitude and more rapidly-varying dynamic in-phase and quadrate signals, namely i(t) and q(t), the signals from the phase sensitive detector are filtered by band-pass filters (BPF) with bandwidths of 0.1 Hz-20 Hz and gains of 51 V/V.

[0231] It should be noted that the sinusoidal voltage signal from the oscillator is amplified through the VCCS and the IA.sub.body stages, which set the mid-band gain, namely A.sub.MB. When determining A.sub.MB, both noise and dynamic range considerations have been taken into account. For improved noise performance, A.sub.MB needs to be sufficiently high. On the other hand, to obtain a large dynamic range that satisfies the linear operation of the circuit, A.sub.MB cannot be arbitrarily large.

[0232] For a typical knee impedance value of R.sub.body=80, high signal-to-noise ratio (SNR) signals can be obtained by setting I.sub.PP2 mA, which is well below the safety threshold. However, it should be noted that Z.sub.Body will vary among different subjects. Therefore, the VCCS is designed as a variable-gain stage that can be tuned to ensure linear operation while not compromising the noise performance. For instance, to obtain high SNR signals from small impedance loads, I.sub.PP can be increased. On the other hand, for measurements from larger impedance loads, to keep the circuit in the linear region, I.sub.PP can be reduced. The maximum and minimum possible values of I.sub.PP are 3.3 mA.sub.pp and 0.6 mA.sub.pp, respectively. The gain of the IA.sub.body, namely A.sub.IA,body, is determined based on the dynamic range constraints. Setting A.sub.IA,body=10.67 V/V, a dynamic range of 300, which exceeds the typical maximum bioimpedance values from the knee, is achieved. The gain of the IA.sub.sense, is set to A.sub.IA,sense=3.47 V/V, which is sufficiently high for generating clocks at the comparator outputs.

[0233] The measurements from the analog front-end are converted to resistance and reactance values by following an automatic calibration procedure. During a measurement period, to compensate for the effects of environmental changes on the measurements, calibration is automatically repeated.

[0234] Calibration

[0235] As with any bioimpedance measurement system, a calibration procedure is necessary to map the static and dynamic voltage signals-namely i(t), q(t), i(t), q(t)-into static and dynamic impedance signals; namely static resistance r(t), static reactance x(t), dynamic resistance r(t) and dynamic reactance x(t). In an ideal synchronous demodulation scheme, i(t) and q(t) would be proportional to r(t) and x(t) respectively, where the impedance being measured is r(t)+jx(t). Therefore, ideally a one-time calibration of the circuitry would be sufficient to map the measured signals to impedance values. However, changes in environmental parameters (e.g., temperature, humidity), and circuit non-idealities, adversely affect the measurement consistency along the course of a measurement. To correct measurement errors caused by changes in the excitation current, I.sub.PP, and therefore increase the robustness of the wearable system, the impedance measurements can be scaled by a correction factor of c.sub.f=2 mA/I.sub.PP (amplitude correction), where I.sub.PP is monitored by A(t). Furthermore, temperature variations also create phase delays at the clocks of the phase-sensitive detection circuitry switches, which affects both magnitude and phase of an impedance measurement. To further rule out those variations, calibration can be performed in an intermittent manner (real-time calibration) by a low-power TI MSP430 series microcontroller.

[0236] As an effort to reduce the calculation burden that real-time calibration could potentially place on the microcontroller, a computationally-efficient two-step calibration procedure is followed: (i) phase correction and (ii) ordinary least squares linear regression.

[0237] The first calibration step is correcting the phase error. The non-ideal switching time, namely t.sub.sw>0, of the demodulator switches in phase-sensitive detection circuitry results in rotation of the measurement vector [i(t) q(t)].sup.T by radians, where =2t.sub.swf.sub.0. Therefore, the phase error is corrected by rotating the vector by radians clockwise to obtain the corrected measurement vector:

[00003]$\begin{array}{cc}\left[\begin{array}{c}\tilde{i}\left(t\right)\\ \tilde{q}\left(t\right)\end{array}\right]=\left[\begin{array}{cc}\cos .Math..Math.& \sin .Math..Math.\\ -\sin .Math..Math..Math.& \cos .Math..Math.\end{array}\right]\left[\begin{array}{c}i\left(t\right)\\ q\left(t\right)\end{array}\right],& \left(3\right)\end{array}$

where (t) and {tilde over (q)}(t) are proportional to r(t) and x(t), respectively.

[0238] The second step aims to map the corrected measurement vector to the impedance vector:

[00004]$\begin{array}{cc}\left[\begin{array}{c}r\left(t\right)\\ x\left(t\right)\end{array}\right]=\left[\begin{array}{c}{m}_R\\ m_X\end{array}\right]\left[\begin{array}{cc}\cos .Math..Math.& \sin .Math..Math.\\ -\sin .Math..Math..Math.& \cos .Math..Math.\end{array}\right]\left[\begin{array}{c}i\left(t\right)\\ q\left(t\right)\end{array}\right]+\left[\begin{array}{c}{c}_R\\ c_X\end{array}\right],& \left(4\right)\end{array}$

where m.sub.R and c.sub.R are the coefficients of linear regression between (t) and r(t). Similarly, m.sub.X and c.sub.X are regression coefficients for the quadrature channel.

[0239] Finding the calibration coefficients in Equation 4namely m.sub.R, c.sub.R, m.sub.X, c.sub.X, and -requires a series of static in-phase and quadrature measurements on test loads of known values. Four series RC test loads, which span the impedance dynamic range, with impedances R.sub.1=23.8, C.sub.1=47 nF; R.sub.2=98.2; R.sub.3=56.4, C.sub.3=94 nF; R.sub.4=75.4, C.sub.4=68 nF are selected. The loads are connected to the analog front-end successively by means of a multiplexer controlled by the microcontroller. The phase correction step is performed by calculating =arctan q.sub.2i.sub.2 using measurements from the purely resistive R.sub.2. Then, linear regression is performed using the measurements from all four test loads to calculate the remaining calibration coefficients. The same coefficients are used to perform the mapping for the dynamic measurements:

[00005]$\begin{array}{cc}\left[\begin{array}{c}.Math..Math.r\left(t\right)\\ .Math..Math.x\left(t\right)\end{array}\right]=\frac{1}{G}\left[\begin{array}{c}{m}_R\\ m_X\end{array}\right]\left[\begin{array}{cc}\cos .Math..Math.& \sin .Math..Math.\\ -\sin .Math..Math..Math.& \cos .Math..Math.\end{array}\right]\left[\begin{array}{c}.Math..Math.i\left(t\right)\\ .Math..Math.q\left(t\right)\end{array}\right],& \left(5\right)\end{array}$

where the factor I/G is introduced to divide out the dynamic channel output gain. The offset vector [c.sub.R c.sub.X].sup.T in Equation 4 is filtered out by the dynamic channel output stages.

[0240] The mapped signals are then used to extract musculoskeletal and cardiovascular features from the knee joints.

[0241] Preprocessing and Feature Extraction Algorithms

[0242] The preprocessing and feature extraction is separately done for static and dynamic signals. The musculoskeletal features of tissue resistance and reactance are extracted from the static signals and the cardiovascular features of heart rate, local pulsatile blood volume and flow rate are extracted from the dynamic signals.

[0243] The preprocessing of the static signals i[n] and q[n] involves conversion to impedance signals r[n] and x[n] using Equation 4. Then, the static signals and the current monitoring signal A[n] are averaged in a 60 second window, to get r.sub.avg, x.sub.avg, and A.sub.avg, respectively. The final joint resistance and reactance measurements of R.sub.measured and X.sub.measured are obtained by performing the amplitude correction explained in the calibration section on r.sub.avg and x.sub.avg using A.sub.avg.

[0244] The signal processing performed on the dynamic signals involves calibration/filtering and ensemble averaging as shown in FIG. 10. FIG. 10 is a physiology-driven algorithm design for extracting cardiovascular parameters automatically from the dynamic component (i.e. IPG) of the bioimpedance signals. First, the signals are pre-processed to reduce noise and detect heartbeats; then heartbeat segmentation is used to enable ensemble averaging to further improve the signal quality; finally, features of blood volume pulse and blood flow are then extracted from these ensemble averaged vector IPG traces.

[0245] The dynamic signals, i[n] and q[n], are first converted to impedance signals via calibration using Equation 5. The dynamic impedance signals are then amplitude corrected using A.sub.avg which is followed by band-pass filtering by an FIR (finite impulse response) filter with bandwidth 0.1 Hz to 20 Hz to obtain [n] and [n]. This step completes the calibration/filtering of the impedance signals before they are ensemble averaged.

[0246] Heartbeats from the knee are detected using i[n]. This signal is first smoothed using a Savitzky-Golay filter of 4.sup.th order and 21 taps, then filtered using a matched filtering approach. The matched filter impulse response (kernel) is a clean, smoothed dynamic in-phase bioimpedance signal, also acquired from the knee previously. This kernel is stored for use and is not adaptive. The matched filtered signal is then differentiated using a Savitzky-Golay Filter of the same kind as used for the smoothing, to form the signal

[00006]$\frac{d}{\mathrm{dt}}.Math.i_m\left[n\right].$

[0247] The signal

[00007]$\frac{d}{\mathrm{dt}}.Math.i_m\left[n\right]$

is fed to a peak detection algorithm to detect heart beats. Peaks in the waveform are searched window-by-window. When a peak is found within the given window, the window size is updated using previous heartbeat intervals. The window is then moved to its next location. To make sure the heartbeats are detected precisely, the window is relocated such that the peak is around the mid-point of the given window. The peak times found are stored in a vector .sub.k. The number of peaks detected per minute gives the heart rate (HR) in bpm.

[0248] The signal i[n] is segmented using the peak times .sub.k. The segments of i[n] (i.sub.k[n]) are stored in each row of the matrix I. The cross correlation of each segment i.sub.k[n], with the kernel is calculated. The maximum of this cross correlation is used to correct the peak times .sub.k such that each i.sub.k[n] is aligned with the kernel. The corrected peak times are stored in the vector .sub.k.

[0249] The peak times .sub.k are used to segment [n] and [n]. The segments [n] and [n] are stored in rows of the matrices and respectively. The ensemble average of [n] is calculated by averaging the segments [n] on a sample by sample basis equation:

r.sub.EA[n]=.sub.k=1.sup.N[n].(6)

[0250] The ensemble average of [n] (x.sub.EA[n]) is then calculated in an analogous manner. The ensemble averaged signals r.sub.EA[n] and x.sub.EA [n] are used for feature extraction. The peak-to-peak amplitudes of these waveforms, r.sub.pp and x.sub.pp, are extracted as they might show differences in injured and healthy knees.

[0251] The signal r.sub.EA[n] is differentiated using the same Savitzky-Golay filter mentioned before, to obtain

[00008]$\frac{d}{\mathrm{dt}}.Math..Math..Math.r_{\mathrm{EA}}\left[n\right].$

On the waveforms, the B, C and X points are identified. The amplitude difference between the points B and

[00009]$C\left(\frac{d}{\mathrm{dt}}.Math..Math..Math.r_{\mathrm{MAX}}.Math..Math.\mathrm{in}.Math..Math..Math.\text{/}.Math.s\right),$

the timing difference between B and X (ejection time, T.sub.ET in s) are used to calculate the local pulsatile blood volume V.sub.blood (in ml) using:

[00010]$\begin{array}{cc}.Math..Math.V_{\mathrm{blood}}={\left(\frac{L}{R_{\mathrm{measured}}}\right)}^2.Math.T_{\mathrm{ER}}.Math.\frac{d}{\mathrm{dt}}.Math..Math..Math.r_{\mathrm{MAX}},& \left(7\right)\end{array}$

where is the resistivity of blood which is taken as 135 cm, L (cm) is the distance between the voltage electrodes and R.sub.measured (), is the measured resistance of the joint. The local blood flow rate Q.sub.local (in ml/min) is calculated using the equation Q.sub.local=V.sub.blood HR. The cardiovascular features HR, V.sub.blood and Q.sub.local will be used in injury assessment as discussed previously.

Results and Discussion

[0252] Circuit Verification

[0253] The designed analog front-end was fabricated on a 64 mm48 mm PCB (FIG. 11(a)). On-board potentiometers were used such that the circuit can also be tuned for other excitation frequencies (7.5 kHz-100 kHz) and current amplitudes (0.6 mA.sub.pp-3.3 mA.sub.pp) as well.

[0254] To verify the calibration procedure, the measurements i.sub.k and q.sub.k were acquired from the four calibration loads mentioned previously using a 3024A oscilloscope (Keysight, Santa Rosa, Calif., US). The calibration parameters were calculated using MATLAB. Static measurements i(t) and q(t) were acquired from the same loads, 10 times within a day.

[0255] From Equation 4, the acquired voltage measurements were converted to impedance measurements using the calculated calibration parameters. The measured versus actual resistance (R) and reactance (X) of the loads are shown in FIG. 11(b). The actual load impedances were measured using an Agilent 34410A 6 digit multimeter. A line was fitted for the R and the X measurements separately (red lines in FIG. 11(b)). The impedance measurements made by the custom analog front-end were linear and consistent within the range 0<R<100 and 0<X<70 (R.sup.21 for both R and X).

[0256] To measure the relative error, four impedances, different then the calibration impedances but within the same impedance range, were used. The measured impedances were compared to those measured using the multimeter. The mean relative measurement errors were 3.9% for R, 5.9% for X, 1.4% for |Z| and 1.8% for Z.

[0257] Connecting a potentiometer to the circuit as a load, the dynamic range of the front-end was tested. The resistance at which the IA measuring across the potentiometer saturates, was defined as the dynamic range and was measured as 345. The measured dynamic range covers the expected knee impedance values varying between 40 and 80 for R and 20 and 10 for X.

[0258] To calculate the noise floor of the measured dynamic impedances, the noise spectral densities of i(t) and q(t) were acquired using an SR785 signal analyzer (Stanford Research Systems, Sunnyvale, Calif., US) with a 0 load across the analog front-end. The cross spectra of i(t) and q(t) were also acquired. The voltage noise spectral densities were mapped to the noise spectral densities of r(t) and x(t) using Equations 4 and 5. The resulting noise spectral densities are shown in FIG. 11(c). The noise floor was calculated for a bandwidth of 0.1-20 Hz. The resulting noise floors were 0.018 m.sub.rms for r(t) and 0.055 m.sub.rms for x(t).

[0259] The current consumption of the front-end was measured as 132 mA when supplied by 5 V regulated from two 9 V batteries, leading to 0.66 W power consumption. The microcontroller power consumption was measured as 150 mW when writing data to the SD card, and 33 mW when sampling with the clock frequency set to 25 MHz. Thus, the overall power consumption for the system was 0.81 W.

[0260] To evaluate the system performance, the electronic and system specifications were compared against numerous bioimpedance measurement systems (see TABLE II). The architectures used for these systems can be categorized as (i) computer assisted designs, (ii) application specific integrated circuits (ASICs), (iii) field programmable gate array (FPGA) designs, and (iv) discrete designs. Computer-assisted systems such as [1-3] cannot be classified as wearable. ASIC systems such as [4-9] are advantageous due to their small size and low power consumption but are expensive to manufacture and limited in terms of programmability. FPGA based systems such as [10, 11] require high current levels, limiting their feasibility for wearable, continuous monitoring applications. Discrete designs such as [12-16] can provide an inexpensive, programmable alternative to existing approaches, can be rapidly prototyped and evaluated in human subjects testing, and-with the design described in this paper-can achieve sufficiently small size and low power consumption for wearable joint health monitoring systems.

TABLE-US-00002 TABLE II Power Noise Cons. (RTI, Measurement Dynamic Reference(s) (W).sup.a m.sub.rms) Error Range () Size Architecture.sup.b Functions [1, 2] |Z|: 0.05 100 < R < Non-wearable, PXI (i) PXI based synth. Static/Dynamic |Z| Z: 0.003 1.1 k based and demod. and Z, spectro. [3] R: 0.02 100 < R < Large, rack (i) Analog front Static |Z| and 1 k mounted modules end + computer Z, spectro. [4] 0.0016 1 < |Z| < 8 cm.sup.2 including (ii) ASIC Static |Z| and (FE) 3.5 k antenna, wearable Z, spectro. 0 < Z < 90 [5] 0.0024 100 5 mm.sup.2 (ii) ASIC Dynamic |Z| (FS) [6] 0.0144 R: 0.2 X: 0.2 R < 54 4.8 3 cm.sup.2 (ii) ASIC Static R and X, (FS) 0.7 < X < spectro. 15 [7] 0.0021 100 < |Z| < 6.95 mm.sup.2 (ii) ASIC Static R and X, multi- (FE) 10 k freq. 0 < Z < 30 [8] 0.0034 R, X: 5% 32 < |Z| < 1.52 mm.sup.2 (ii) ASIC Static R and X, multi- (FS) 5.3 k freq. [9] 58 438 * 7 7 mm.sup.2 (ii) ASIC Static R (FE) 345 (FS) [10] 3.7 5 R < 3.2 k 136 145 mm.sup.2 (iii) FPGA based Static/Dynamic |Z| (FS) and Z, spectro. [11] |Z|: 1.2% |Z| > 2 k (iii) FPGA based Static |Z| and Z Z: 0.18 [12] |Z|: 1.1% Z: 1% R > 10 k (iv) AD5933 Static |Z| and controlled by C Z, spectro. [13] R: 1% 50 < R < Wearable (iv) AD5933 Static R and X, 1.6 k based spectro. [14] |Z| < 1 k Portable (iv) Discrete with Static |Z|, Dynamic sync. demod. |Z| [15] |Z|: 0.36 9 < |Z| < Handheld, 0.5 kg (iv) AD8302 Static |Z| and Z: 0.0049 5.7 k controlled by C Z, spectro. 0 < Z < 180 [16] 0.546 R: 0.4 X: 0.3 145 40 4 mm.sup.3, 30 g (iv) AD5933 Static R and X, (FS) based spectro. Present 0.66 0.018 R: 3.9% X: 5.9% |Z| < 345 64 48 mm.sup.2 (FE) (iv) Discrete with, Static and Dynamic R, Invention (FE) |Z|: 1.4%, sync. demod. Static and Dynamic X 0.81 Z: 1.8% (FS) .sup.aFS = Full system, FE = Front-end only; Noise RTI in [5] was calculated for a bandwidth of 20 Hz using the noise spectral density reported. .sup.bArchitecture sorted into four groups: (i) Computer-assisted architecture; (ii) ASIC; (iii) FPGA; (iv) Discrete design.

[0261] TABLE II is a comparison of electrical and system specifications for EBI circuits and systems sorted by architecture. The references include: [0262] [1] B. Sanchez, J. Schoukens, R. Bragos, and G. Vandersteen, Novel Estimation of the Electrical Bioimpedance Using the Local Polynomial Method. Application to In Vivo Real-Time Myocardium Tissue Impedance Characterization During the Cardiac Cycle, Biomedical Engineering, IEEE Transactions on, vol. 58, pp. 3376-3385, 2011. [0263] [2] B. Sanchez, E. Louarroudi, E. Jorge, J. Cinca, R. Bragos, and R. Pintelon, A new measuring and identification approach for time-varying bioimpedance using multisine electrical impedance spectroscopy, Physiological Measurement, vol. 34, p. 339, 2013. [0264] [3] A. Hartov, R. A. Mazzarese, F. R. Reiss, T. E. Kerner, K. S. Osterman, D. B. Williams, and K. D. Paulsen, A multichannel continuously selectable multifrequency electrical impedance spectroscopy measurement system, Biomedical Engineering, IEEE Transactions on, vol. 47, pp. 49-58, 2000. [0265] [4] J. Ramos, J. L. Ausin, A. M. Lorido, F. Redondo, and J. F. Duque-Carrillo, A wireless, compact, and scalable bioimpedance measurement system for energy-efficient multichannel body sensor solutions, Journal of Physics: Conference Series, vol. 434, p. 012016, 2013. [0266] [5] L. Yan, J. Bae, S. Lee, T. Roh, K. Song, and H.-J. Yoo, A 3.9 mW 25-Electrode Reconfigured Sensor for Wearable Cardiac Monitoring System, Solid-State Circuits, IEEE Journal of, vol. 46, pp. 353-364, 2011. [0267] [6] S. Lee, S. Polito, C. Agell, S. Mitra, R. F. Yazicioglu, J. Riistama, J. Habetha, and J. Penders, A Low-power and Compact-sized Wearable Bio-impedance Monitor with Wireless Connectivity, Journal of Physics: Conference Series, vol. 434, p. 012013, 2013. [0268] [7] A. Yufera, A. Rueda, J. M. Munoz, R. Doldan, G. Leger, and E. O. Rodriguez-Villegas, A tissue impedance measurement chip for myocardial ischemia detection, Circuits and Systems I: Regular Papers, IEEE Transactions on, vol. 52, pp. 2620-2628, 2005. [0269] [8] P. Kassanos, L. Constantinou, I. F. Triantis, and A. Demosthenous, An Integrated Analog Readout for Multi-Frequency Bioimpedance Measurements, Sensors Journal, IEEE, vol. 14, pp. 2792-2800, 2014. [0270] [9] N. Van Helleputte, M. Konijnenburg, J. Pettine, J. Dong-Woo, K. Hyejung, A. Morgado, R. Van Wegberg, T. Torfs, R. Mohan, A. Breeschoten, H. de Groot, C. Van Hoof, and R. F. Yazicioglu, A 345 uW Multi-Sensor Biomedical SoC With Bio-Impedance, 3-Channel ECG, Motion Artifact Reduction, and Integrated DSP, Solid-State Circuits, IEEE Journal of, vol. 50, pp. 230-244, 2015. [0271] [10] S. Kaufmann, A. Malhotra, G. Ardelt, and M. Ryschka, A high accuracy broadband measurement system for time resolved complex bioimpedance measurements, Physiological Measurement, vol. 35, p. 1163, 2014. [0272] [11] S. Sun, L. Xu, Z. Cao, H. Zhou, and W. Yang, A high-speed electrical impedance measurement circuit based on information-filtering demodulation, Measurement Science and Technology, vol. 25, p. 075010, 2014. [0273] [12] C. Margo, J. Katrib, M. Nadi, and A. Rouane, A four-electrode low frequency impedance spectroscopy measurement system using the AD5933 measurement chip, Physiological Measurement, vol. 34, p. 391, 2013.

[0274] [13] F. Seoane, J. Ferreira, J. J. Sanchez, and R. Brags, An analog front-end enables electrical impedance spectroscopy system on-chip for biomedical applications, Physiological Measurement, vol. 29, p. 5267, 2008. [0275] [14] L.-Y. Shyu, C.-Y. Chiang, C.-P. Liu, and W.-C. Hu, Portable impedance cardiography system for real-time noninvasive cardiac output measurement, Journal of Medical and Biological Engineering, vol. 20, pp. 193-202, 2000. [0276] [15] Y. Yang, J. Wang, G. Yu, F. Niu, and P. He, Design and preliminary evaluation of a portable device for the measurement of bioimpedance spectroscopy, Physiological Measurement, vol. 27, p. 1293, 2006. [0277] [16] T. Schlebusch, Ro, x, thlingsho, x, L. fer, K. Saim, Ko, x, M. ny, and S. Leonhardt, On the Road to a Textile Integrated Bioimpedance Early Warning System for Lung Edema, in Body Sensor Networks (BSN), 2010 International Conference on, 2010, pp. 302-307.

[0278] As seen from TABLE II, the present invention has higher resolution in impedance measurements than all the similar conventional systems shown. This enables the present invention to sense blood-flow-related impedance changes from the knee. This high measurement resolution is achieved in an energy-efficient manner with a small footprint. Furthermore, compared with other systems in TABLE II, the present invention is the only one (that can be deployed in a wearable device) enabling real-time calibration to minimize measurement errors due to environmental changes.

[0279] Physiological Measurement Results and Discussion

[0280] Human subject testing was performed on nine subjects; seven control subjects with no recent history of injuries to the knees and two injured subjects with recent unilateral knee injuries (ACL or meniscal tear). The studies were approved by the Georgia Institute of Technology IRB as well as the AHRPO. The circuit was interfaced to the body via Ag/AgCl gel electrodes, positioned as shown in FIG. 12(a). The proximal current electrode was placed three inches above the crease on the quadriceps tendon, towards the inner side of the knee. The distal current electrode was placed three inches below the crease on the popliteal fossa, towards the outer side of the knee. The voltage electrodes were placed adjacent to the current electrodes. To evaluate the ensemble averaging algorithm, ECG signals from the control group were acquired using a BioNomadix (Biopac Systems Inc, Goleta, Calif., US) wireless ECG acquisition module.

[0281] All signals acquired were recorded on Biopac data acquisition hardware and processed on MATLAB. For the human subject studies, a one-time calibration followed by amplitude correction was performed.

[0282] During the testing protocol, each subject sat upright with his/her back resting against the wall and legs extended forward. While the subject was still, 90 seconds of the signals i(t), q(t), i(t), q(t), A(t) and the ECG (for the control group) were acquired from both knees separately. The data between 20 s and 80 s were processed to rule out any motion artifacts at the beginning and the end of the measurement cycle (from initial positioning of the body). For pulsatile blood volume calculations, the distance between the voltage electrodes on each knee were measured.

[0283] The resulting static impedances for each knee from each subject are plotted in FIG. 12(b), and the differences between the magnitude of resistance and reactance in left versus right knee are shown in FIG. 13. The left versus the right knee impedances of nearly all control subjects were within 12 from each other in the R and 3.2 from each other in the X axes. The difference between the resistance and reactance of the left and the right knees had no statistical significance for the control group. However, even for the small feasibility study sample size of two injured subjects, there was a significant difference (p<0.01) in the impedances of the involved and contralateral knee compared to the control group. The injured side on both subjects had lower resistance, R, by an average of 13.7 and negative reactance, X, by an average of 5.25. These results are consistent with physiological expectations (reduced resistance due to increased edema, reduced negative reactance associated with damaged cell membranes).

[0284] The pre-processing and feature extraction algorithms described were used to calculate the cardiovascular features HR, V.sub.blood and Q.sub.local for the control group. The signals obtained at intermediate and final steps of the ensemble averaging algorithm are presented in FIG. 12(c). The accurate detection of heart beats enabled the dynamic (IPG) impedance signals to be segmented accurately. Therefore, high SNR waveforms were obtained after ensemble averaging leading to clear detection of B, C, and X points.

[0285] These features were also calculated using an ECG-assisted ensemble averaging algorithm for comparison. It was observed that the cardiovascular features calculated with and without the ECG were consistent (FIG. 12(d)) with no statistically significant differences observed. Furthermore, no significant differences were observed in the cardiovascular features on the injured side compared to the contralateral side.

[0286] For one of the control subjects, a separate experiment was performed to examine the effects of vasoconstriction on the dynamic impedance signal measured from one knee. Specifically, the methodology was analogous to a cold pressor test, with the subject's bare foot being submerged in ice water while the IPG signal was measured across the knee (not submerged) to increase downstream peripheral vascular resistance (PVR). The purpose of this experiment was to evaluate the sensitivity with which the system developed in this work could detect minute changes in knee joint blood flow associated with modified downstream PVR.

[0287] The subject's skin temperature was measured on the bare foot prior to submerging and found to be 30 C. The impedance signals were acquired from the knee for 60 seconds while the subject was seated, with the tested leg extended and resting on a support. The foot of the same side was immersed into ice water until the foot skin temperature dropped to 17 C. The foot was taken out of the cold water and impedance signals were again recorded for 60 seconds from the knee in the same position as before.

[0288] A plot of the resistance signal ensemble averaged using the algorithm described above (r.sub.EA(t)) and its derivative

[00011]$\left(\frac{d}{\mathrm{dt}}.Math..Math..Math.r_{\mathrm{EA}}\left(t\right)\right)$

taken using a Savitzky-Golay filter are shown in FIG. 12(e). The minute changes in this IPG signal associated with downstream vasoconstriction can be observed on both waveforms as a significant drop in amplitude (20 m decrease for the resistance, and 0.2/s decrease for the time derivative of resistance). Accordingly, the system was sufficiently sensitive to detect both the reduction in local blood volume pulse and blood flow with increased PVR, consistent with physiological expectations.

Assessing Joint Physiology or Structure in A Manner Relevant to Assessing Overuse of The Joint

[0289] In an exemplary embodiment related to assessing overuse of the joint, vector EBI measurements are taken from the joint using a tetrapolar electrode configuration before, after, and/or at different intervals during, the training session or competition. Changes in the resistive and/or reactive components of the EBI measurements are quantified. Simultaneously obtained inertial measures such as joint angle and body posture/position are taken to ensure that the physical conditions of the user during the EBI measurements were the same for each usage. These changes in EBI characteristics are then outputted to the user as a score indicative of joint swelling/tissue damage.

[0290] In another exemplary embodiment, joint acoustical emission measurements are taken using contact microphones (e.g., piezoelectric based acceleration or velocity sensors) placed around the joint. Inertial measures are simultaneously obtained such as joint angle, velocity, and/or position of the limbs or limb segments. The user performs several prescribed movements while wearing the sensors, and the acoustical emissions are measured in the context of particular activities. These measurements are obtained before, after, and/or at various intervals during the training session or competition. Algorithms, such as graph mining in conjunction with graph community factor quantification, are used to derive quantitative characteristics from the acoustical emission signals for each usage. Changes in these characteristics throughout the training session or competition are quantified and outputted to the user as a score indicative of joint inflammation/structural disruption.

[0291] In another exemplary embodiment, vector EBI, joint acoustical emission, and inertial measures are obtained simultaneously from the joint before, after, and/or at different intervals during, the training session or competition. Changes in the characteristics of all three measures are extracted and fused together to provide a single score indicative of overall joint health. This score is then outputted to the user.

[0292] In another exemplary embodiment, the EBI measurements are taken together with local measurements of skin temperature such that the changes in skin temperature are used to correct the EBI characteristics, since changes in skin temperature can modulate the filtration/absorption of fluid from the capillaries into the interstitial space and thus confound the measurement of local swelling.

[0293] In another exemplary embodiment, the joint acoustical emissions are measured with a combination of electret or microelectromechanical systems (MEMS) microphones (which capture airborne sounds) and contact microphones. The two measurement modalities are fused, for example by determining which signal has more interference and extracting characteristics only from the other signal. Since microphones capturing airborne sounds are more susceptible to background noise than contact microphones, while contact microphones are more susceptible to rubbing noise from contact with skin or clothing, the interferences are distinctly different and thus the fusion results in improved signal quality overall.

[0294] In another exemplary embodiment, the sensors are attached to the elbow and/or shoulder of a user's throwing arm and contralateral arm to compare the sensed parameters on both sides such that the characteristics on one side versus the other can be quantified. Thus, the user's contralateral arm is used as a control, and a more sensitive and accurate joint health or use score can be outputted to the user to direct further training, competition, or rehabilitation regimens.

[0295] In another exemplary embodiment, sensor data can be obtained from the joint before and after treatments, such as in the training room. For example, data can be obtained before and after ice is applied to the joint, to assess the differential effects of the treatment on the joint and improve the sensitivity and accuracy with which the joint health or use score derived from the sensor data.

[0296] In an exemplary embodiment, the sensor data can be processed locally on the device using a microprocessor such that the indication can be provided directly to the user.

[0297] In another exemplary embodiment, the sensor data can be transmitted wirelessly to a phone, tablet, laptop, or other computing device for remotely processing the data and providing an indication to the user.

[0298] In another exemplary embodiment, the sensor data can be used to quantify inflammation non-invasively (without the need for obtaining blood from the subject) in tracking progression after muscle injury resulting from eccentric loading. For example, the characteristics of the joint acoustic emission and bioimpedance measurements can be fused for serial measurements taken periodicallyor continuouslythroughout training/eccentric loading exercises.

[0299] FIG. 14 illustrates the concept of acquiring two sets of recordings with the system, the first taken as a baseline recording, and the second taken following a perturbation. The first recording can include a combination of multiple modalities of sensors, and the second recording would typically include the same combination of sensors. The perturbation can comprise any set of activities performed by the user (e.g., exercise), changes to the environment or joint conditions enforced externally (e.g., changing skin temperature or joint stiffness conditions), or a combination of both (e.g., running on a treadmill while the stiffness is being changed locally). The perturbation can be applied artificially for the purposes of enhancing the information extracted for measurement purposes or can be an activity that the user was planning to perform anyway (e.g., activities performed by an athlete during a normal practice or training workout). The data taken following the perturbation (in the second recording) can then be compared against the data taken during the baseline recording using a variety of algorithms, such as clustering analysis and graph mining techniques.

[0300] Whether the characteristics of acoustical emissions from the joint change in a quantifiable and monotonic manner in response to increased vertical loading forces for a standardized movement is herein evaluated. It is assumed that increased vertical loading forces will lead to more interactions between internal surfaces of the knee and thus will increase the complexity of the emitted sounds. The reason for this variability in the signal could be the increased number of surfaces interacting, and increased friction between surfaces causing more complex interactions and ensuing acoustic emissions.

[0301] FIG. 15 provides an illustration of the assumption, and FIG. 16 shows the measurement setup used for assessing loading effects, based on a vertical leg press with varying weight and the measurement of acoustical emissions from the knee.

[0302] Graph mining algorithms were leveraged to quantify this complexity and evaluated the approach in a study of able-bodied subjects. This is believed the first-time the effects of joint loading forces on acoustical emissions is quantified. Quantifying loading based on wearable measurements of joint acoustical emissions throughout activities of daily living and in a broader range of movements and exercises presents continued avenues for research.

Methodology

[0303] Human Subjects Study and Measurement Protocol

[0304] Twelve healthy subjects with no prior injuries were recruited for the study which was approved by the Georgia Institute of Technology Institutional Review Board (IRB). Subject demographics are described in TABLE III.

TABLE-US-00003 TABLE III DEMOGRAPHIC DATA FOR STUDY PARTICIPANTS Male Female Number of Subjects 9 3 Age (mean , in years) 24.3 1.9 25.3 2.1 Height (mean , in cm) 175.1 3.5 155 2.7 Weight (mean , in kg) 74.3 9.9 50.7 7.2

[0305] For each subject, four miniature (7.95.54.1 mm) contact microphones (BU-23173-000, Knowles Electronics LLC., USA) were attached to the medial and lateral sides of the patella and superficial to the lateral and medial meniscus using Kinesio Tex tape (see FIG. 16). These microphone, which are piezoelectric sensors with broad bandwidth (>20 kHz) and low output noise (7 Vrms A weighted), were selected as they provide high quality acoustical pick-up from the body while maintaining a small footprint amenable for wearable use. The subjects were then asked to perform ten repetitions of vertical leg press with different loading conditions. Vertical leg press was selected as the exercise for this study since it allows for vertical loading at the knee joint to be varied in a controlled manner up to a sufficiently large load (i.e. body weight) to notice effects on the acoustical emissions. There was a total of four loading conditions, starting with zero load (no weights), and increasing by a third of the subject's body weight up to the full weight of the subject. The audio signals from each microphone were pre-amplified using a custom analog front-end comprising a voltage regulator, setting a 3V supply used for powering the microphones and amplifiers, a low noise amplifier with a voltage gain of 100, and 180 kHz bandwidth, and a bandpass filter (bandwidth: 16 Hz-20 kHz, Butterworth). The amplified signals were then sampled at 50 kHz (16 bits/sample) using NI USB-6225 data acquisition hardware (NI, Austin, Tex., USA). All signals were recorded on a laptop using LabVIEW System Design Software (NI, Austin, Tex., USA) and were processed using MATLAB (The Mathworks, Naticks, Mass., USA) and RStudio (RStudio, Boston, Mass., USA).

[0306] Joint Sound Pre-Processing and Feature Extraction

[0307] The overall signal processing and feature extraction steps are illustrated in FIG. 18. The acoustical signals were acquired during ten repetitions of vertical leg press cycles on the dominant leg of the subject. The signals were digitally filtered using a Kaiser-window finite impulse response bandpass filter with a bandwidth of 400 Hz-20 kHz to remove the low-frequency interface noise between the tape and the microphone. The filtered signals were standardized with zero mean and unity variance to balance out any variations in the signal amplitude among sensors that could result from differences in contact pressure against the skin.

[0308] The normalized signals were then divided into segments (windows) with a duration of 200 ms and 90% overlap between successive segments. This segment duration and the overlap allowed multiple joint sound signatures to be present within a given frame. The features derived from each segment are summarized in TABLE IV. The features were selected empirically and are commonly used in other audio signal processing applications. For example, the mel-frequency cepstrum coefficient (MFCC) is prevalent in speech recognition analysis, for discriminating speech, music, and background noise.

[0309] From each windowed segment of each microphone, a total of 69 features were extracted and stored in a matrix. For each loading condition, the features were vertically concatenated extracted from all four microphones into one single matrix. Based on common statistical learning rules, the number of features acquired for the dataset is a reasonable value.

TABLE-US-00004 TABLE IV AUDIO FEATURES FOR KNEE JOINT SOUNDS Feature Name General Description Significance Energy Total signal energy Short term energy is expected to exhibit high variation over successive speech frame Zero Crossing Rate Rate of sign changes Exhibit higher values in the case of noisy signals (i.e. noisier if there is higher loading forced on the knee joints) Energy Entropy Measure of abrupt changes in the Low value in abrupt energy changes (i.e. energy level low peaks if there are higher loads affecting the joints) Spectral Centroid and Spread Center of gravity of spectrum/ Higher values correspond to brighter second central moment of sounds (i.e. brighter sounds if higher spectrum loads are endured on the joints) Spectral Entropy Like entropy but in frequency Higher value in sounds for more loads domain on the joints Spectral Flux Measure of spectral change Lower value if the signals are more between two successive frames consistent (i.e. lower the values will be as more loads affect the joint sounds) Spectral Rolloff Frequency below which 90% of Higher value for wider spectrum (i.e. the signal energy (magnitude) is higher in joint sounds with more loads concentrated are enforced) MFCCs Coefficients that make up a First 13 MFCCs carry enough representation of the short-term discriminative information to compare power spectrum of a sound, based joint sounds with different loading on a linear cosine transform of a conditions log power spectrum on a nonlinear mel scale of frequency. Band Powers Power of the signal in 29 distinct Higher frequency band powers will frequency bands, between 30 exhibit high values at frames where joint logarithmically spaced sounds are most abrupt. (i.e. joints frequencies in the range of 1 kHz- affected by more loads will have higher 15 kHz values) Chroma Vector, Fundamental Features widely used in music- These features could indicate which Frequency, and Harmonic related applications segments contain high values (i.e. joint Ratio sounds affected by higher loads) Standard deviation, mean, Basic features for signal Higher values for noisy signals (i.e. median, minimum, maximum processing noisier if there is higher loading enforced on the joints)

[0310] Proof-of-Concept Study of Loading Effects during Walking

[0311] An alternative exercise was considered using an Alter-G (AlterG, Inc., Fermont Calif., USA), an anti-gravity treadmill that assists patients with fast recovery. Illustrations for the measurement setup and the exercise movement are depicted in FIGS. 19-20. The microphone placements were the same as the locations for the vertical leg press. As a proof-of-concept, one subject was recruited, approved by the IRB, to perform 20 seconds of walking on the treadmill with two different weight conditions: 20% of the body weight (minimum), and 100% of the body weight (maximum). The same signal processing and feature extraction methods were used.

[0312] Graph Mining Algorithm

[0313] It was believed that increasing the vertical loading forces on the knee would increase heterogeneity among acoustical signals captured by the microphones. To investigate this, the concept in graph theory of quantifying heterogeneity by locating, and computing the number of, communities within the graph were used. The combined features from microphones for a single loading condition as a data matrix X was considered.

[0314] It is assumed that more loads should increase the number of sound sources in the knee and this would increase the heterogeneity within the captured signals by different microphones. Accordingly, the distribution of X should be modeled. Although this can be done using some statistical models such as Gaussian or student t-distributions, these techniques require strong assumptions about the high-dimensional shape of the data (e.g., ellipsoid versus convex) and model parameters (e.g. mean and standard deviation) which can cause many problems, such as unreliable bandwidth estimation for applied kernel density function.

[0315] To overcome these challenges, a k-Nearest Neighbor graph (kNN) from graph theory was reconstructed that previously has been successfully used by researchers to model and cluster high dimensional bioinformatics data. The constructed graph from a knee which experienced smaller loads should be less heterogeneous than the one that experiences higher loads. This heterogeneity can be modeled with the number of complex communities in the related graph: a greater number of graph communities should be needed to describe sounds emitted from a knee which is loaded with higher forces.

[0316] In this work, a kNN graph is defined for each dataset X Let KG={V, E} indicate the kNN graph corresponding to X where V={v.sub.1, v.sub.2, . . . , v.sub.N} is the set of vertices and E.Math.VV represents the set of edges among v.sub.i. In this graph, each vertex v.sub.i indicates one row (acoustical window) in X. To model the local neighbor of each window x.sub.i in X, the corresponding vertex v.sub.i is connected to its k nearest neighbors using Euclidean distance. In this work, k was chosen to be 10 empirically. Other values were also investigated (e.g. 5-15) and similar results were obtained.

[0317] If only the Euclidean distance values are considered to assign related weights of edges between v.sub.i with its nearest neighbors, noisy data points would engender many problems. If there are some v.sub.is expanding the dispersed zones between two different communities, it may not distinguish these two communities and merge them as one single community incorrectly. Hence, weights are reassigned to each graph edge using dice similarity, such that we incorporate the properties of each point's neighborhood rather than relying on Euclidean distance alone in attributing points to clusters or communities. The dice similarity of v.sub.i and v.sub.j means twice the number of common neighbors divided by the sum of the degrees of v.sub.i and v.sub.j. Assuming v.sub.i and v.sub.j indicate two connected vertices within the kNN graph, the assigned weight for the edge between these two vertices is defined as:

[00012]$\begin{array}{cc}{w}_{\mathrm{ij}}=\frac{2*.Math.\left\{A_i.Math.B_j\right\}.Math.}{(.Math.D_i.Math.+.Math.D_j).Math.},& \left(8\right)\end{array}$

where A.sub.i and B.sub.j denotes the set of the neighbors of v.sub.i and v.sub.j, respectively. Also, the degree of v.sub.i and v.sub.j are represented as D.sub.i and D.sub.j, respectively and finally, the notation |*| is the number of elements in a set.

[0318] Once the weighted kNN graph is extracted, a community detection algorithm is applied to extract all the potential communities (clusters) within the kNN graph. There are several community detection algorithms studied in graph theory that could be applied. In this work, the Infomap community detection algorithm is employed to quantify the communities of the kNN graph, since Infomap has been applied successfully in various areas of graph mining in different fields such as bioinformatics.

[0319] Infomap uses the probability flow of random walks on the network as a proxy for information flows and clusters the graph into multiple communities. The algorithm searches for a partitioning of the kNN graph to minimize the expected description length of a random walk and seeks to compress the description of information flow visited by a random walker on the network. Using Huffman code, all v.sub.i visited by a walker are recorded and coded. The walker takes a reasonable amount of time within the same community which results in longer walking process. The computational complexity of this algorithm is approximately O(|E|). In this work, the number of detected communities is shown with GCF (Graph Community Factor) which represents the heterogeneity of extracted kNN graph from the data matrix X

[0320] One important note is that discovering of the potential communities in the kNN graph is tantamount to finding the number of clusters (dense areas) in a high dimensional dataset X. Applying regular clustering methodologies such as K-means and Gaussian Mixture Models are not possible in this problem, as these methods require the knowledge of the number of clusters (dense populations of acoustical windows) within the data matrix. It is further noted that applying a kernel-based density clustering algorithm (as it automatically estimates the number of dense areas in data) on a 69-dimensional dataset X to find the clusters is challenging and is not practical. The difficulty is that the curse of dimensionality causes the density detection in high dimensions (in this problem 69) to be very time consuming and statistically not robust.

Results and Discussion

[0321] Changes in the GCF with Loading for All Microphones

[0322] The use of the graph mining algorithm was evaluated to quantify the changes of acoustical emissions from the knee joints with respect to different vertical loading forces on twelve subjects. Four contact microphones were used to collect the joint sounds from various locations on the knee (medial and lateral of patella and meniscus). FIG. 21 illustrates the increasing trend of heterogeneity and the calculated GCF value with respect to different loading conditions for one representative subject. The knee graph is constructed from the individual loading data matrix where each data point (vertex in the graph) represents all the time and frequency domain features for one windowed frame of the acoustical emissions and different colors characterize different communities. The GCF values for no loading, one-third body weight (BW), two-thirds BW, and BW were 26, 34, 38, 40, respectively.

[0323] FIG. 22 provides a boxplot of the GCF values calculated for all subjects for each loading condition. The mean values for no load, one-third BW, two-thirds BW, and BW were 30, 32, 36, and 39. Since the sample size is not large, the non-parametric paired Kolmogorov-Smirnov test was used to calculate the p-value (p<0.01). For twelve subjects, we demonstrated the increase in GCF of the acoustic emissions from the joints with respect to the increasing load level on the knee.

[0324] Changes in GCF Across Microphone Locations

[0325] The characteristics of the acoustical signals were also evaluated across four different microphone locations to determine which locations had the most heterogeneity. Each microphone data matrix consists of all the segments and the features for the four loading conditions. FIG. 23 shows the boxplot of average GCF value for all subjects with respect to the microphone locations (numbers). Microphones 1,2,3, and 4 have mean GCF values of 26.4, 32.6, 33.5, and 23.8 respectively.

[0326] The same non-parametric paired Kolmogorov-Smirnov test was used to calculate the p-value (p<0.01). Referring to FIG. 17, the results showed that locations 2 and 3 (medial side of the patella and superficial to the lateral meniscus, respectively) seem to show higher variation than the microphones placed on 1 and 4 (lateral side of the patella and superficial to the medial meniscus, respectively). The underlying anatomical sources of this variation are hypothesized as follows: microphone 1 is located on the superior lateral aspect of the knee which is principally superficial to the femur, quadriceps tendons, muscle, and fat. Microphone 3 placement includes the tibia collateral ligament, muscles, and the semi-membranous bursae, a jelly filled sacs in between ligament and bone, along with the previously mentioned lateral meniscus and associated connective tissue. In addition, the fibula connects to the tibia on this side of the leg, which could cause considerable differences from the sounds produced on the opposite side. The fibular collateral ligaments are located on the lateral side, close to microphone 3, which could contribute to more heterogeneity in that location since it is another source of tension on the joint when load is applied. Locations 1 to 4 (lower heterogeneity) and locations 2 to 3 (higher heterogeneity) makes an X through the knee. There are several anatomical structures that follow a similar course through the region, and as such may be particularly potent contributors to the variations and heterogeneity seen. These structures include the two proximal heads of the gastrocnemius, the popliteus muscle/ligaments, and the anterior cruciate ligament (ACL) and posterior cruciate ligament (PCL).

[0327] Changes in GCF during Walking with Loading

[0328] Using the graph mining technique from the acoustical signal obtained from one subject, it could be evaluated that the GCF value increased from 28 to 36 as the body weight changed from 20% (minimum) to 100% (maximum), respectively (see FIG. 24). This showed that changing the loading on the knee while walking could have similar impact on the knee joint acoustical emissions as for the vertical leg press exercise.

Wearable Sensor System to Evaluate and Manage Throwing Arm Health

[0329] In yet a further embodiment, developing wearable technology that can quantify significant changes in joint (shoulder and elbow) health status for pitchers that could inform sports medicine practitioners and aid in the prevention of overuse injuries could be beneficial. Recovery training and rehabilitation decisions (e.g., continued need for modalities and/or recovery pitch counts or velocities) before the next hard practice or game can be made with quantitative, objective, joint health data complementing current approaches.

[0330] The collection of key joint health parameters using our present sound, swelling, and activity context measurements can be developed and validated on the knee joint, and be translated to the elbow and shoulder.

[0331] FIG. 25 illustrates a timeline for a pitcher's training/competition schedule, and how physiological-data-driven recovery schedules can potentially be used to avoid overuse injury. Major League Baseball (MLB) has an interest in preventing overuse injuries and managing recovery from arm (elbow/shoulder) injuries in pitchers. Current approaches employ self-reported discomfort/pain, physical examination and pitch count metrics, which are either subjective or based on performance indicators and not biomarkers. Quantitative biomarkers sensitive to worsening or improving subclinical and clinical injuries are needed to manage throwing arm health in pitchers.

[0332] FIGS. 26(a)-(d) illustrate an overview of the present wearable sensor system to evaluate and manage throwing arm health.

[0333] A wearable physiological monitoring system to assess knee health that incorporates measures of acoustical emissions (i.e., joint sounds), swelling based on electrical bioimpedance (EBI), and joint angle and kinematics with advanced machine-learning algorithms for an individualized joint health score is disclosed herein. Acoustics capture information regarding alignments of articulating surfaces while EBI can detect extremely small differences in tissue edema due to heating and/or cooling the limb. These have been applied these measures to discriminate healthy from injured knees and quantify post-surgery recovery. This same approach potentially can be applied to the arm (elbow and shoulder) to provide quantitative biomarkers of injury and recovery during directed movements in pitchers.

[0334] It is assumed that an elbow and shoulder health scorederived from sounds, swelling, and activity context metricswill detect subtle biological changes, which can then be used to manage recovery in pitchers and thus reduce the potential risk of arm injuries. A decrement in passive glenohumeral external rotation (lying supine) was over two times more likely for pitchers to be placed on the disabled list with a throwing arm injury.

[0335] Here, the measurement of joint angles achieved during an active throwing motion are coupled simultaneously with joint sounds during the motion. Swelling in the joint picked up by electrical bioimpedance in the present sensor system might also explain decreased range of motion that occurs intermittently throughout the season. Collectively, these signals can be used as a gauge to evaluate each pitcher's individual response to treatment or previous pitching practices as an objective recovery signal. The sensing modalities that are quantifying represent the same aspects (sounds, swelling, and activity context) that are qualitatively evaluated in physical exams, and thus should provide sufficiently rich information to more comprehensively assess changing throwing arm health.

[0336] One first obtains measurements of acoustical emissions, EBI, and activity context (inertial measurement) data from pitchers to optimize the location of the sensors, and the types of movements, to facilitate the highest quality of data acquisition. The present invention includes methods that determine what constitutes a high-quality signal for each of these modalities. For example, for measurement of acoustical emissions, we have developed algorithms for quantifying the signal to interference and noise ratio (SNIR), which can then be compared for different measurement locations for the sensors. Similarly, for the EBI measurements, measurement repeatability and variability in the resistance and reactance measurements across the recording time have been used as an index of quality.

[0337] It is assumed that the locations across the elbow will be analogous to placements used at the knee, since many of the characteristics of this hinge jointsuch as range of movement, internal bone structure, etc.are similar. It is assumed that the shoulder might require more in-depth studies and envision requiring minor changes to the sensor packaging and adhesion approaches. In terms of activities used to generate acoustical emissions, it is anticipated that the activities for the elbow to closely mirror our approach for the kneeunloaded elbow flexion/extension exercises will be used, combined with pronation/supination which is an important motion for pitching and putting spin on the ball. For the shoulder, movements that are typically used in the physical exam currently will be used to evaluate crepitus such as internal and external rotation, shoulder flexion, and shoulder internal rotation in 90 degrees of abduction (90-90).

[0338] After optimizing sensor placement, packaging, and activities for the elbow and shoulder, measurements of acoustical emissions, EBI, and joint angle baseball players, for example (5 pitchers, and 5 infielders), will be obtained. The measurements will be obtained at three time points throughout the season (to assess cumulative wear-and-tear on the joints). At each time point, sensor data will be measured (a) before, (b) immediately after, (c) 24 hours after, and (d) 48 hours after an exhaustive pitching session (to assess acute effects of a fatiguing workout on the joints, and delayed onset inflammatory effects such as edema).

[0339] Signals from both arms will be measured, such that the contralateral side can provide a control for each subject that can then be used to normalize inter-subject differences in the data. If any of the pitchers are injured during the season, we will take recordings at two-week intervals throughout the rehabilitation period following the injury. This will allow the data analytics efforts to both focus on pre-injury (prediction) and post-injury (rehabilitation status) quantification of joint health/risk.

[0340] The data analytics efforts will then have two focuses: (1) using unsupervised machine learning (i.e., graph mining) to facilitate clustering of the measured data to determine which signals, and features of signals, provide the best capability in detecting fatigue and wear-and-tear (cumulative throughout the season, incorporating metrics such as the number of pitches throughout the season) for the joints, and (2) using supervised machine learning (i.e., support vector machines) to provide preliminary approaches for quantifying the risk of acute overuse injury throughout the season by mapping sensor data to physical exam, symptoms, and other reference standard clinical data regarding the status of the players. These reference standard measures can include player-reported symptoms/pain, athletic trainer evaluations, and functional surveys.

[0341] Evaluation will focus on assessing the following: (1) quantifying whether high-quality measurements are obtained in throwing and non-throwing arm, (2) the differences in the pre- and post-workout measures, (3) the differences across the short-term recovery phase following a typical, large-volume pitching effort, and (4) the differences throughout the course of a season. To quantify whether the high-quality measurements are obtained, first the EBI signals will be visualized and expect the resistance to be below 150 and negative reactance to be below 50 at a single frequency of 50 kHz. The cole-cole characteristic in the spectroscopy plots will also be reviewed, as these clearly will indicate a good signal quality.

[0342] In terms of the joint sounds, first the SNIR will be computed to determine the optimal location for the microphone placement. Then a plot of the signals and detect clicks that are in high amplitudes and in short duration will be made, and they will be analyzed in the spectroscopy to look for high energy and broad bandwidth. These high amplitude clicks could correlate to the popping or crunching sounds which may indicate the friction between the joints or soft tissue impingement. In terms of evaluating the differences in the pre- and post-workout, it is anticipated the EBI parameters at a single frequency (50 kHz) to slightly decrease and the absolute difference of these values between the throwing and the non-throwing arm to increase as there would be a minor effect of edema and a structural wear-and-tear.

[0343] As for the immediate recovery phase (24-48 hours) after a large pitching effort, it is hypothesized that the absolute differences between the two arms would eventually return to the normal state (i.e. pre-workout levels) over time (although the exact time course may differ for each individual and each session). For this longitudinal testing, it is predicted that throughout recovery, EBI parameters would decrease with variations to occur with an inflammatory flare up (potentially signaling inadequate recovery or overuse syndrome). As for the elbow and shoulder sounds, the frequency domain features and custom features related specific to the clicks will be extracted and analyzed, and then any changes that occur will be observed using an unsupervised machine learning algorithm (i.e. increase in the graph heterogeneity).

CONCLUSIONS

[0344] The present invention incorporates a discrete design for high resolution EBI measurements from the knee joint based on embedded systems concepts. The combined use of high performance analog front-end electronics and digital programmability using a microcontroller allow high quality static (slowly varying over the course of hours to days) and dynamic (rapidly varying on the order of milli-seconds) impedance measurements on a platform that is appropriate for a wearable device.

[0345] The overall system was designed and demonstrated from end-to-end, including the design of the circuit, customized physiology-driven algorithms for detecting features from the measured signals, and human subject experiments to both evaluate the capabilities in detecting small changes in local bioimpedance due to edema and modified blood flow. The invention includes the mechanical packaging to encapsulate the electronics in a wearable sleeve around the knee joint and collect data serially during the recovery from acute knee joint injury. The present wearable system based on the engineering foundation herein presented enables high resolution, quantitative assessment of both the structural and hemodynamic characteristics of the knee joint longitudinally for the first time, paving the way to better understanding joint recovery physiology, and designing closed-loop personalized therapies to accelerate the recovery process.

[0346] Further, a method of using a graph mining algorithm to quantify the impact of loading on knee joint sounds is established. With increasing loading conditions, the acoustical emissions became more heterogeneous. Furthermore, there were more variations in microphone placement at the medial side of the patella and the lateral side of the meniscus.

[0347] Numerous characteristics and advantages have been set forth in the foregoing description, together with details of structure and function. While the invention has been disclosed in several forms, it will be apparent to those skilled in the art that many modifications, additions, and deletions, especially in matters of shape, size, and arrangement of parts, can be made therein without departing from the spirit and scope of the invention and its equivalents as set forth in the following claims. Therefore, other modifications or embodiments as may be suggested by the teachings herein are particularly reserved as they fall within the breadth and scope of the claims here appended.