HEADSET FOR DIAGNOSIS OF CONCUSSION

Abstract

A system and method for detecting brain concussion includes detecting and measuring of acceleration at one or more points on a subject's head. Sensors, which can be accelerometers placed against the head, detect and measure natural motions of the patient's head due to blood flow in the brain and resultant movement of tissue in the brain. An observation is then made, as compared with data corresponding to non-concussion, of a change in frequency response pattern exhibited when accelerations are plotted as a function of time or frequency, to identify probable concussion. Preferably the observation and comparison are made by a computer using an algorithm.

Claims

1. A system for detecting brain concussion in a human patient, the system including a neural network and a headset, the headset comprising: a three-axis accelerometer configured to sense skull motion produced by pulsatile cerebral blood flow; an adjustable snap-on headband connected to a housing and configured to place the three-axis accelerometer temporally on the head of the patient; a digitizer in the housing to digitize the signal from the three-axis accelerometer; a rechargeable battery in the housing to power the headset; a data storage medium for recording the digitized signal from the three-axis accelerometer; and a USB connection to charge the rechargeable battery and download the recorded digitized signal from the data storage medium to a laptop; wherein the headset is configured to transmit recorded digitized signal to the neural network trained by inputting digitized signals from the three-axis accelerometers collected from a plurality of subjects with concussion or with non-concussion, which is adapted to identify and categorize unique features producing specific signatures of the signals received from the three-axis accelerometers, and which compares a frequency response pattern derived from the recorded digitized signal with frequency response data corresponding to non-concussion and identify probable concussion in the patient based on differences between the frequency response pattern recorded from the head of the patient and the frequency response data corresponding to non-concussion.

2. The system of claim 1 further comprising a wireless transceiver for transmitting the recorded digitized signal to a mobile device.

3. The system of claim 1, wherein the headset positions the three-axis accelerometer against a temple of the patient such that the accelerometer detects and measures natural motions of the patient's head due to blood flow in the brain and resultant movement of tissue in the brain.

4. The system of claim 1, further comprising another three-axis accelerometer.

5. The system of claim 1, further comprising another three-axis accelerometer, and wherein the headset positions each three-axis accelerometer against a different temple of the patient such that the accelerometer detects and measures natural motions of the patient's head due to blood flow in the brain and resultant movement of tissue in the brain.

6. The system of claim 1, further comprising a heartbeat sensor.

7. A headset for detecting brain concussion in a human patient, the system comprising: a three-axis accelerometer configured to sense skull motion produced by pulsatile cerebral blood flow; an adjustable snap-on headband connected to a housing and configured to place the three-axis accelerometer temporally on the head of the patient; a digitizer in the housing to digitize the signal from the three-axis accelerometer; a rechargeable battery in the housing to power the headset; a data storage medium for recording the digitized signal from the three-axis accelerometer; and a USB connection to charge the rechargeable battery and download the recorded digitized signal from the data storage medium to a laptop; wherein the headset is configured to transmit recorded digitized signal to a laptop which applies a diagnostic algorithm to the digitized signal from the three axis accelerometer to identify probable concussion in the patient, the diagnostic algorithm utilizing a neural network generated by inputting digitized signals from the three-axis accelerometers collected from a plurality of subjects with concussion or with non-concussion, the neural network being adapted to identify and categorize unique features producing specific signatures of the signals received from the three-axis accelerometers.

8. (canceled)

9. The headset of claim 7 further comprising a wireless transceiver for transmitting the recorded digitized signal to a mobile device.

10. The headset of claim 7, further comprising another three-axis accelerometer.

11. The headset of claim 7, further comprising another three-axis accelerometer, and wherein the headset positions each three-axis accelerometer against a different temple of the patient such that the accelerometer detects and measures natural motions of the patient's head due to blood flow in the brain and resultant movement of tissue in the brain.

12. The headset of claim 7, further comprising a heartbeat sensor.

13. A method for detecting brain concussion in a human patient, the method comprising: providing a headset comprising an adjustable headband connected to a housing and a three-axis accelerometer configured to sense skull motion produced by pulsatile cerebral blood flow, the adjustable headband configured to place the three-axis accelerometer temporally on the head of the patient, a digitizer in the housing to digitize the signal from the three-axis accelerometer, a rechargeable battery in the housing to power the headset, a data storage medium for recording the digitized signal from the three-axis accelerometer, a USB connection to charge the rechargeable battery and download the recorded digitized signal from the data storage medium to a laptop; placing the headset on the head of the subject such that the three-axis accelerometer detects and measures natural motions of the patient's head due to blood flow in the brain and resultant movement of tissue in the brain; transmitting the digitized signal from the three-axis accelerometer to a laptop; generating a diagnostic algorithm by inputting digitized signals from the three-axis accelerometers collected from a plurality of subjects with concussion or with non-concussion into a neural network, the neural network being adapted to identify and categorize unique features producing specific signatures of the signals received from the three-axis accelerometers; and applying the diagnostic algorithm to the digitized signal from the three axis accelerometer to identify unique features producing specific signatures in the digitized signal, which are indicative of probable concussion in the patient.

14. (canceled)

15. The method of claim 13 wherein the headset further comprises a wireless transceiver for transmitting the recorded digitized signal to a mobile device.

16. The method of claim 13, wherein the headset positions the three-axis accelerometer against a temple of the patient such that the accelerometer detects and measures natural motions of the patient's head due to blood flow in the brain and resultant movement of tissue in the brain.

17. The method of claim 13 wherein the headset further comprises another three-axis accelerometer.

18. The method of claim 13 wherein the headset further comprises another three-axis accelerometer, and wherein the headset positions each three-axis accelerometer against a different temple of the patient such that the accelerometer detects and measures natural motions of the patient's head due to blood flow in the brain and resultant movement of tissue in the brain.

19. The method of claim 13 wherein the headset further comprises a heartbeat sensor.

20. The headset of claim 7, wherein the headset positions the three-axis accelerometer against a temple of the patient such that the accelerometer detects and measures natural motions of the patient's head due to blood flow in the brain and resultant movement of tissue in the brain.

Description

DESCRIPTION OF THE DRAWINGS

[0021] FIG. 1A is a view of a concussion headset on a head of a patient according to an embodiment of the invention.

[0022] FIG. 1B a block diagram showing components of the concussion headset system according to an embodiment of the invention.

[0023] FIGS. 1C and 1D are frontal and side views of a concussion headset according to an embodiment according to an embodiment of the invention.

[0024] FIGS. 2A and 2B are frontal and side views schematically indicating accelerometer sensors located on a patient's head according to another embodiment of the invention.

[0025] FIG. 2C is a block diagram showing components of the system and method of the invention.

[0026] FIG. 3A is a flow chart showing an embodiment of the process of the invention.

[0027] FIG. 3B is a flow chart showing an embodiment of the process of the invention.

[0028] FIG. 4A is a chart indicating Phase 1 testing of subjects using the system of the invention.

[0029] FIG. 4B is a chart indicating Phase 2 testing of subjects.

[0030] FIG. 5 is a frequency plot of a particular subject, showing a pattern indicating non-concussion.

[0031] FIGS. 6A to 6D are FFT plots of four different subjects, all indicating concussion.

[0032] FIG. 7 is a chart of the R.sub.1 factor of the algorithm of the invention against time.

[0033] FIGS. 8A-8C are charts plotting R1 and R2 factors for several different populations of recordings using the system of the invention.

[0034] FIG. 9 shows an example of a Campbell diagram, as an indication of concussion.

[0035] FIGS. 10 and 11 show examples of time-domain waveforms for typical non-concussed and concussed subjects.

DESCRIPTION OF PREFERRED EMBODIMENTS

[0036] Preferred embodiments of the invention are explained below in terms of system components and tests that have been performed. Initial discussion is in regard to testing performed to identify a specific pattern indication of concussion using cranial accelerometry (phase 1), to test the identified pattern against blinded data to verify its efficacy in detecting concussion (phase 2).

Accelerometry Measurement Equipment

[0037] Concussion can be detected using a single temporally-placed sensor. FIG. 1A is a view of a concussion headset on a head of a patient according to an embodiment of the invention. As shown in FIG. 1A, the concussion headset 100 includes a snap-on headset design which includes an adjustable band 114, connected to a housing 112. The adjustable band 114 secures the headset to the head 102 of the patient and places an accelerometer 110 in contact with the head of the patient. As shown in FIG. 1A, accelerometer 110 is placed temporally on the head of the 102 of the patient. The exterior of housing 112 includes a switch 134, LED indicator 132 and USB port 126. In an alternative embodiment two accelerometers 110 are provided in the headset 100—one on each temple of the patient (bitemporally). In an alternative embodiment, the headset may also be provided with a heartbeat sensor such as an earlobe or in-ear Photoplethysmography (PPG) sensor. Such heartbeat data can be used to augment and/or synchronize the accelerometry data.

[0038] FIG. 1B a block diagram showing components of the concussion headset system according to an embodiment of the invention. As shown in FIG. 1B, headset 100, includes accelerometer 110 coupled to a digitizer 120. Digitizer 120 digitizes the analog signal from the accelerometer. The headset 100 includes data storage 122 for storing several recordings before upload. The headset is powered by a rechargeable battery 130. A Bluetooth transceiver 128 provides a Bluetooth connection 162 to download data to a mobile device 160. The USB port 126 provides a connection via a USB cable 152 to a laptop 150 to charge the rechargeable battery 130 and also download recordings from data storage 122 to the laptop 150 for analysis. The components of headset 100 are controlled by controller 124. FIGS. 1C and 1D are detailed frontal and side views of an embodiment of concussion headset 100 according to an embodiment according to an embodiment of the invention.

[0039] The concussion headset system collects and stores data derived by sensing small motions of the surface of the head of the subject produced by pulsatile cerebral blood flow and its impact on the skull. This is accomplished by accelerometer 110 that is positioned within the adjustable headset 100 to contact the scalp temporally. Transduced accelerometer data from accelerometer 110 is digitized with the digitizer 120 and recordings are stored in data storage 122. The recordings can be downloaded from data storage 122 to the laptop 150 for analysis as described below. Acceleration of the skull is detected and measured using the accelerometer, and these measurements are sent to a computer. The data include frequency and intensity of vibrations of the skull at the location of the accelerometer. The data from the accelerometer are analyzed and investigated. If the data shows a change in frequency content of skull motion in selected frequency ranges as compared with data corresponding to non-concussion, this is taken to indicate probable concussion.

[0040] For example, in an embodiment, accelerometer 110 is a three axis accelerometer. A three axis accelerometer comprises three single axis accelerometers oriented orthogonally to each other. An example of a three axis accelerometer is a 3-Axis digital accelerometer module available from the TESSEL PROJECT. The 3-Axis digital accelerometer module utilizes an MMA8452Q, 3-axis, 12-bit/8-bit digital accelerometer device available from NXP SEMICONDUCTORS. This smart, low-power accelerometer device incorporates three capacitive, micro-machined transducers oriented on three axes. The device also incorporates an analog to digital converter (see digitizer 120 of FIG. 1B) which converts analog outputs from the three capacitive transducers into 12-bit digital accelerometry data. The device provides three channels of digital accelerometry data (one channel per axis) at 12 bits of resolution per channel.

[0041] FIGS. 2A and 2B are frontal and side views schematically indicating accelerometer sensors located on a patient's head according to another embodiment of the invention. This device comprises a headset 8 with six accelerometers 10, as indicated in FIGS. 2A-2B, as well as a data collector unit 12 with amplifier 14 and digitizer 16, a detachable coaxial cable 18, and laptop computer 20, all shown schematically in FIG. 2C.

[0042] The system collects and stores data derived by sensing small motions produced by pulsatile cerebral blood flow and its impact on the skull. This is accomplished by six highly sensitive accelerometers 10 that are positioned within an adjustable headset 8 to contact the scalp bitemporally, bifrontally, occipitally, and at the skull vertex. Note that concussion can be detected using only a single temporally-placed sensor, although two opposed temporal sensors are preferred. The accelerometers are very sensitive, typically 500 mV/g or more. Such an accelerometer is a model from Dytran Instruments, Inc. of California. A heart rate sensor 13 is also provided using a PPG (photoplethysmography). An omnidirectional sound pressure level (SPL) sensor 11 is positioned on the top of the headset for the purpose of ambient noise eradication (see FIGS. 2A and 2B). In an alternative embodiment, three axis accelerometers, as shown in FIG. 1B, may be used at one or more positions in the headset.

[0043] Transduced accelerometer data are digitized with the digital signal processor 12 and sampled with the laptop computer 20 (see FIG. 2C). The system detects and analyzes low-frequency signals from the accelerometers, below 500 Hz and for most analysis no higher than 100 Hz. Much analysis can be at 30 Hz or less. Correlated noise patterns between the SPL sensor and accelerometer data may be removed during processing.

Accelerometry Measurement Methodology

[0044] FIG. 3A is a simplified flow chart outlining the method of the invention. The block 25 indicates detection and recording of acceleration at one or more selected points on the head of a subject. This data is digitized and, in a preferred embodiment, a fast Fourier transform (FFT) is performed on the data, as indicated at 26. A waveform is produced, per the block 28, as a plot of energy or intensity versus frequency. This waveform is compared to a reference waveform (block 30), to determine whether concussion is indicated. The dependent blocks 32, 34 and 36 indicate that this comparison can be done visually, by algorithm, or, as explained below, via a Campbell diagram. If comparison is done visually, the FFT waveform will be shown on a computer monitor and can be compared to a typical non-concussion waveform or to a typical concussion waveform, or both. A mathematical algorithm developed according to the invention is explained below, as is the Campbell diagram. The block 38 indicates the system provides a conclusion as to whether concussion is indicated, and the probability of accuracy can accompany this conclusion.

[0045] FIG. 3B is a flow chart outlining the method of the invention for developing and using a diagnostic algorithm. A diagnostic algorithm in an initial phase 360. Subsequently the diagnostic algorithm is used for diagnosis of concussion in the field 370 with a patient whose condition is unknown. Initially cohorts of concussion subjects 300 and non-concussion subjects 310 are enrolled. The headset for which a diagnostic algorithm is to be developed in then used to collect accelerometry data. See, e.g. FIG. 4B. The block 25 indicates detection and recording of acceleration at one or more selected points on the head of a subject using a device under test. The accelerometry data also includes the data from the other sensors in the system such as from a heart beat sensor (PPG) and/or sound pressure level sensor if present.

[0046] All of the collected data is digitized and, in a preferred embodiment, a fast Fourier transform (FFT) is performed on the data. The digitized data from the cohorts of concussion subjects 300 and non-concussion subjects 310 (training set) is then analyzed at step 320 using one or more of machine learning, mathematic analysis, expert system, neural net system, and/or artificial intelligence system. The analysis is used to develop a differential diagnostic algorithm 330. The analysis reveals one or more putative diagnostic algorithm. The putative diagnostic algorithm is then verified against a blinded set of accelerometry data from a second cohort of concussion subjects and second cohort of non-concussion subjects (validation set). See, e.g. FIG. 4B. Steps 320, 330 and 340 may be repeated until a valid and predictive diagnostic algorithm is defined for accelerometry data collected using the headset under test.

[0047] The diagnostic algorithm 350 can then be applied to a field diagnostic system. For example, the diagnostic algorithm may be provided to computer 20 of FIG. 2C, laptop 150 of FIG. 1B or Mobile device 160 of FIG. 1B. Alternatively, if the diagnostic algorithm is too computationally intensive for such devices, the accelerometry data may be forwarded by such data to a remote or cloud computing system for analysis. In the field, accelerometry data is collected from a patient 350 (of unknown status) using the headset. The accelerometry data is analyzed using the diagnostic algorithm 350. The system then calculates concussion of non-concussion (with probability) for the patient 350. The block 38 indicates the system provides a conclusion as to whether concussion is indicated, and the probability of accuracy can accompany this conclusion.

[0048] In addition to concussion, cranial accelerometry data may be used to non-invasively diagnose a number of abnormal cranial conditions which cause micro or macro changes in the pulsatile blood flow with the brain or brain tissues. Such conditions include but are not limited to: stroke; ischemic stroke; vessel occlusion, hemorrhagic stroke; aneurysm; stenosis; peripheral bleeding; thrombotic, embolic, lacunar and hypo-perfusion strokes; intra-cranial hemorrhage; subarachnoid hemorrhage. Different diagnostic algorithms can be developed to allow a headset to diagnose these conditions. In order to develop such diagnostic algorithms different cohorts of patients with and without the condition must be assessed with the headset to provide the training data set and verification data set for development and verification of the diagnostic algorithm. Once developed the diagnostic algorithm and headset may be used for assessment of patients in the field suspected of having the particular condition.

Example: Neural Net

[0049] A neural net is one example of a machine learning, expert system or artificial intelligence system. For example, the headsets disclosed herein can be used in a system and method for detecting a vascular condition non-invasively in the human body, by a collection of signal information from local, small regions of the vasculature. This is accomplished in a preferred embodiment by attaching, or contacting an array of accelerometers, or other sensors in a headset, to the head of subjects and recording vibration signals. The vibration signatures of blood vessel structures such as branches, aneurysms, stenosis or other structures using random, periodic, band limited or transient analysis provide a library for further processing. The library becomes part of a cerebral modeling arithmetic computer using basis functions, or other artificial neural network.

[0050] Neural networks typically are trained by inputting a set of known inputs and known outputs and allowing the weights of the neural connections to change to optimize the matching of the inputs and outputs. When a new input is presented to the neural network the output closest to the input is given the highest output even if the input is not a perfect match to any of the training set. The data to build the neural network (NN) and/or data store for correlation are obtained by acquiring data from patients with known pathologies as well as controls, and optionally with studies of constructed or cadaver vasculatures. The array of different accelerometer signatures produced by features within the vasculature is quite large due to the variable physiology of patients but there are generally only a few major underlying features. This type of data is ideal for using neural networks to identify the features causing the signature to be categorized. As the library of unique features and the associated signatures grows the neural network will improve in correctly identifying features in patients. The data are formatted according to the requirements of each processing method before being used for training or algorithm construction. The developed signature library and/or a diagnostic algorithm derived therefrom, is then used to diagnose conditions with the results presented to the physician in a clinically relevant manner.

Accelerometry Measurement Testing

[0051] FIG. 4A is a chart indicating Phase 1 testing of subjects using the system of the invention. FIG. 4B is a chart indicating Phase 2 testing of subjects. As explained in greater detail below, phase 1 was the process whereby accelerometry data indicative of a concussion pattern were determined, and phase 2 was evaluation of these findings against a blinded set of accelerometry data. In all testing, subjects had accelerometry measurements and concurrent two-lead electrocardiograms. In players with a concussion, multiple sequential measurements were obtained. Sport Concussion Assessment Tool 2 was used to assist clinical determination of concussion. The following explanation pertains to methods used to acquire data for both phases 1 and 2.

Sport Concussion Assessment Tool

[0052] Sport Concussion Assessment Tool 2 (SCAT2) was documented and accompanied all accelerometry recordings except those obtained postseason. In players with a concussion, multiple sequential measurements were obtained. Sport Concussion Assessment Tool 2 was used to assist clinical determination of concussion. Sport Concussion Assessment Tool 2 is a standardized tool introduced at the 2008 third Consensus Conference on Concussion in Sport held in Zurich. All scores from sections of the SCAT2 tool contribute to a possible total of 100 points. Scoring fewer points in any section lowers the aggregate score. For instance, baseline scores for high school athletes averaged 88.3+6.8. In our study, declines in SCAT2 scores were used to assist in the clinical determination of whether a head impact resulted in a concussion. SCAT2 or SCAT3 or similar neurocognitive testing can be used for other indications of concussion such as accidents.

Subjects and Sample Size

[0053] Candidates for this observational study included all football players (grades 9-12) at a northern California high school during the 2011 football season.

[0054] Visual inspection of data from concussed and non-concussed subjects revealed that in the non-concussed subjects, cranial motion had a series of peaks that typically declined (diminished in amplitude) with increased frequency. FIG. 5 shows the frequency components plot obtained from the cranial accelerometry recording of a typical non-concussed subject. It is notable that the fifth and sixth harmonic peaks have lower amplitude than do the first to third harmonic peaks. This was a consistent finding in subjects without concussion.

[0055] In the subjects with concussion, several additional peaks, representing higher frequencies, appeared. These new peaks arose at frequencies including and between 5 and 12 times the heart rate frequency. Three high-frequency ranges were used in algorithm development to differentiate between concussed and non-concussed subjects. FIGS. 6A-6D show the frequency component plots of four concussed subjects. Note that the fifth and sixth harmonic peaks have higher amplitude than do the first to third harmonic peaks. This increase in higher frequency harmonic content continues in paired bands at about 5 to 6 times the heart rate, 8 to 9 times the heart rate and 11 to 12 times the heart rate. This increase in higher frequency harmonic content can also be observed in the time-domain waveforms as shown in FIGS. 10 and 11, described below.

[0056] Several algorithms are described to quantify the differences between concussed and non-concussed subjects. Other algorithms are also possible. The algorithm can be in several forms, but primarily it will detect, as noted above, an increase in energy of skull motion in frequency ranges above about the fourth harmonic of the patient's heartbeat. In a mathematical algorithm these data need to be “normalized”, and in one form of algorithm the averaged energy of, for example, the fifth and sixth harmonics of the heartbeat is compared against (i.e. divided by) a maximum value of one or more or an average of several of the lower harmonics, below the fourth harmonic. This ratio, which is often called herein R.sub.1, is compared against a selected threshold. Such a threshold, in one embodiment of the invention, is set at 1.0, but different thresholds can be selected, based on desired levels of sensitivity and specificity. The higher the threshold for R.sub.1 is set, the lower the sensitivity but the higher the specificity, and vice versa.

[0057] Preferably another ratio factor is also included in the described algorithm, which can be called R.sub.2. R.sub.2 represents “normalized” data from the eighth and ninth harmonic peaks. The average value of those peaks is also divided by a maximum or average of one or more of the lower harmonics, the same denominator used for R.sub.1. In one embodiment of the algorithm the R.sub.2 threshold is set at 0.66. Thus, concussion is indicated if R.sub.1≥1.0 and R.sub.2≥0.66. Concussion is contra-indicated in the case where both R.sub.1 and R.sub.2 are below the set thresholds.

[0058] One such algorithm defines up to three factors (R.sub.1, R.sub.2 and R.sub.3) based on the relative height of the peaks, with the higher frequencies compared with the lower. The lower three harmonic values (first, second and third) define the denominator for these three variables. R.sub.1 is defined as the average of harmonics 5 and 6 amplitude divided by the maximum of the lower three harmonic amplitudes. R.sub.2 is the average of harmonics 8 and 9 amplitude divided by the maximum of the lower three harmonic amplitudes. R.sub.3 is the average of harmonics 11 and 12 amplitude divided by the maximum of the lower three harmonic amplitudes. Higher values of R.sub.1 through R.sub.3 indicate higher energy within the higher frequencies of skull motion. R.sub.1 and R.sub.2 can be used to create a Boolean value (true or false). This Boolean value is true if R.sub.1>1.0 AND R.sub.2>0.66; a Boolean value of “true” defines concussion. R.sub.1 and R.sub.2 (or either of them) can also be used as a continuous variable that follows the clinical time course of concussion. Likewise R.sub.3 can used to discriminate between concussed and non-concussed subjects by its increase over non-concussed subjects' values and by comparing it to R.sub.1 and R.sub.2. In this case the values of R.sub.1, R.sub.2 and R.sub.3 should diminish with increasing harmonic number.

[0059] The sensitivity and specificity of the described algorithm in detecting concussion from a set of subjects playing American football are shown in Tables 1 and 2.

TABLE-US-00001 TABLE 1 Sensitivity in Detecting Concussion Clinically Concussed C NC Sensitivity, CI, Subjects* (True+) (False−) % % 13 10 3 76.9 46-94 *Thirteen subjects were confirmed by SCAT2 and clinical assessment to have suffered a concussion. The concussion algorithm pattern was seen in 10 of these subjects and was not seen in 3 of these subjects. C, concussion; NC, no concussion; CI, confidence interval (P = 0.95).

TABLE-US-00002 TABLE 2 Specificity in Detecting Concussion No. C NC Specificity, CI, Recordings (False+) (True−) % % Baseline and 18 5 13 72.2 46-89 postseason recordings* Baseline 15 0 15 100.0  74-100 recordings** Baseline 58 7 51 87.9 76-94 recording*** Total 91 12 79 87.0 78-93 recordings *Baseline and postseason recordings from subjects (n = 9) with no reported or suspected concussion for the duration of study. **Baseline recordings from subjects (n = 15) who, at some time after baseline recording, reported a possible concussion and had multiple recordings related to that event. Thirteen of these 15 subjects were confirmed to have a concussion by SCAT2 and clinical assessment (Table 1). ***Baseline recordings from subjects with no reported or suspected concussion at the time of recording. C, concussion; NC, no concussion; CI, confidence interval (P = 0.95).

[0060] An important aspect of the invention is the ability to detect concussion early. Typically peak data exceeding thresholds, in the procedure described above, are not exhibited for a few days after a concussive event, e.g. about day 4. The factors R.sub.1 and R.sub.2 may not cross the thresholds until day 3 or day 4, or even longer in some cases. It is known through cognitive concussion testing that some delay occurs in concussion manifesting itself (although maximum indications in cognitive testing tend to occur earlier than those from the subject algorithm). However, the system of the invention can detect the rise of these factors toward concussion even in the first day or two after an event of trauma to the head. This is accomplished by observing the velocity of change in the patient's R.sub.1 and R.sub.2 values in the period leading up to the concussion symptoms exhibiting themselves, referred to herein as the period of “developing” concussion, typically about four days. With the system of the invention, the data can be observed as showing clear and definitive movement toward crossing the thresholds for R.sub.1 and R.sub.2, in a patient with concussion. Thus, days before the data analysis shows positive for concussion, the velocity of movement in that direction will indicate concussion. There previously existed no other reliable means for detecting concussion in this “developing” period.

[0061] After R.sub.1 and/or R.sub.2 reach peak values, i.e. in the case of R.sub.1 the maximization of the frequency-domain peaks at about the fifth and sixth harmonics of the patient's heartbeat, a period of usually several weeks of declining values ensues, called the “recovery” period.

[0062] To gather insight into the time course of the presence of the concussion pattern, we plotted factor R.sub.1 over time in subjects who were able to provide multiple recordings after concussion. FIG. 7 is an example of R.sub.1 plotted against time for one subject with a concussion. The first cranial accelerometry recording was the day following the suspected concussive event (SCE). R.sub.1 generally varied smoothly and followed the clinical course of concussion. However, in subjects with multiple recordings, the concussion pattern returned to a normal pattern at a time beyond self-reporting of symptoms as measured by neurocognitive testing (catalog and assessment of symptoms). This is notable in that concussion patients tend to be cleared by neurocognitive to return to sports activities too early. Analysis indicates that the time course of this pathophysiological change begins immediately after concussion and outlasts the duration of clinical concussion before it eventually resolves. Using the method described above to detect concussion leads to delays of 0 to 7 days from injury to detection. Another factor, R.sub.3, the rate of change of R.sub.1 or R.sub.2 with time (R.sub.1 or R.sub.2 velocity) may be used to predict a concussion within hours of an injury. A given individual, when measured repeatedly, has a cluster of R.sub.1 and R.sub.2 which may be similar to that as shown in FIG. 8A, at lower left. When a recording provides an R.sub.1 and R.sub.2 that differs by an amount greater than the average variation in typical subjects, then that change provides a measure of concussion detection. The more recordings and factors obtained, the more confident the diagnosis of concussion. These factors could be obtained from continuous recordings to further improve the detection speed.

[0063] Likewise a visual examination of the time domain or frequency domain plots of the average brain pulse or a visual examination of both domains of the average brain pulse provides a rapid method of detecting concussion, as explained below with reference to FIGS. 10 and 11. These changes can be further displayed in another visual algorithm using color intensity waterfall (CIWF) plots. In this type of plot, the overlap FFT of the subject recordings are plotted in a horizontal line with frequency on the horizontal line axis. The intensity of each frequency bin of the FFT is color-coded; a typical color-coding uses blue for the lowest value and red as the highest value following the color spectrum for values in between. Each subsequent recording FFT is plotted below the previous recording so that the vertical axis is time. The convention is that time increases toward the plot origin. With this type of plot, shifting harmonic content is observed as increasing color intensity over time in the 5 to 13 harmonic frequencies.

[0064] The method of the invention also encompasses additional techniques to optimize the identification of peaks on the plot of accelerometer data in the frequency domain, using algorithms. As discussed above, one preferred algorithm developed pursuant to the invention has been to look at the values at the fifth and sixth harmonics as a basis for R.sub.1, and also at the eighth and ninth harmonics as a basis for R.sub.2. When an algorithm is applied based on the harmonics, the peaks will seldom fall precisely at the harmonic frequencies being investigated. In spite of this, the algorithm functions very well, with high sensitivity and specificity. Improvement can be made, however, by more closely identifying the actual frequency locations of the peaks that tend to occur around the fifth and sixth harmonics and also around the eighth and ninth harmonics. This can be done visually on a computer monitor, and it could also be done using a pattern recognition program, similar to those used for facial pattern recognition. Techniques are also available to identify these peaks accurately using an algorithm that does not use the visual appearance of a time domain or frequency domain chart. As one example, the algorithm can include a maximizing or optimizing feature by which, after the data are analyzed at the fifth and sixth harmonics and the eighth and ninth harmonics (as well as at lower harmonics as discussed above), the frequency under examination can be shifted up or down for each of the fifth, sixth, eighth and ninth harmonics to find the peak value within a selected frequency band of the nominal harmonic. By this method the actual peaks of the four approximate harmonics that are used to calculate R.sub.1 and R.sub.2 can be pinpointed, for more accurate analysis.

[0065] Another related algorithm to detect concussion is a variation of CIWF, the Campbell diagram. Another important aspect of the invention is the use of a Campbell diagram as an indicator of concussion, or as a confirmation. The responses of a concussion patient's brain to vascular pulsing are frequency dependent, and the Campbell diagram was developed for frequency dependent functions, such as turbine design in jet engines. Vibration response in a turbine tends to be different at different revolution speeds. Since concussed brain responses are also frequency dependent, i.e. heartbeat rate dependent, a typical waterfall diagram of the gathered data on a patient will usually not produce sharp lines—a patient's heartbeat rate can vary with time. However, the data can be plotted to produce sharper lines, as eigenfrequency lines, if heart rate is represented on the vertical axis and frequency on the horizontal axis. The eigenfrequency of a concussion patient typically are essentially radial lines emanating and fanning out from the theoretical point 0,0. In particular, if the “hot” color bands (red, orange, indicating high intensity) follow such radial, fanning lines, this indicates the harmonics of the system are changing in frequency with the driving function. That is to say, as the speed decreases, then the harmonics as an ensemble decrease in a pattern of the fan going down to zero. If, on the other hand, the structure is not responding to the driving force, i.e. to the frequency, but is simply a structural response, then the bands will be vertical. Therefore, one can use these bands to detect whether this is a structural change. It appears that in the normal brain (without concussion), the bands tend to be vertical; changing the heart rate does not shift the peaks of the harmonics. With concussion, changing the heart rate does change the position of the harmonics such that they follow the eigenfrequency lines down to zero. This is another method of detecting concussion and potentially detecting it much earlier than the R.sub.1 or R.sub.2 or velocity can do. The Campbell diagram provides an efficient reference that can be used as a primary determination for concussion indication (or not), or which can be used as a check against the conclusion reached via another algorithm such as that discussed above. The harmonic peak locations can be compared to the eigenfrequency lines by a correlation function to determine how well the structure responds to the varying heart rate.

[0066] The Campbell diagram was developed to detect damaging harmonic resonances in rotating machinery such as jet turbine engines. In human subjects the heart rate varies with respiration, providing a natural variation to the brain pulse, so that the Campbell diagram is based on heart rate. Each of the individual heart rate brain pulse recordings is sorted. The CIWF plot is modified so that the vertical axis is no longer time but heart rate with the highest heart rate FFT line plots farther from the plot origin and the lowest heart rate FFT line plots closer to the plot origin (see FIG. 9). Lines, called eigenfrequency lines, are drawn from each harmonic on the highest frequency FFT line plot, with each eigenfrequency line connecting to the waterfall plot origin. If the changes in the brain pulse are due to a concussion, the subsequently lower heart rate frequency FFT line plots will have lower harmonic peak locations, aligning with each corresponding eigenfrequency line, and higher heart rate FFT plots will have higher harmonic peak locations. Correlation of the harmonic peak location to the eigenfrequency lines, fanning out essentially from the plot's origin, provides a factor for concussion detection. This fanning out Campbell diagram seems not to occur without concussion.

[0067] The precise pathophysiology of concussion remains undefined, but likely is related in some way to injury that includes blood flow abnormalities, fluid accumulation, and/or structural changes. A shift of harmonic energy toward higher frequencies (or an increase at the higher frequencies) is likely caused by an increase in brain resonance from the force induced by pulsatile blood flow entering the skull. This may be caused by a stiffer brain and could be accounted for by brain edema or perhaps changes in cellular structure. However, brain edema has not been demonstrated in persons with clinical concussion, so this explanation is by no means certain. It is possible that disruption of cerebral autoregulation might produce phase changes in brain resonance and result in higher harmonics of the cardiac-induced forcing function into the closed skull space.

[0068] It was not part of our study to determine which subjects continued to suffer from concussion at the end of the season as determined by SCAT2 test results. We did not obtain SCAT2 test results from subjects at the end of the season, so cannot conclude who did or did not continue to register a concussion by that test method. Throughout the study, we used clinical judgment to determine the presence or absence of concussion. It is a recognized limitation of concussion diagnosis that clinical judgment and psychometric testing are subject to variability. However, this clinical judgment could not have been biased by knowledge of accelerometry data because these data were kept blinded until they had been fully collected and analyzed.

[0069] The observation that in certain subjects the concussion pattern returned to the non-concussion pattern after reporting of symptoms resolution but then returned to the concussion pattern after exercising suggests that return to activity is perhaps premature if it is not based on an objective physiological determination.

[0070] FIGS. 8A, 8B and 8C are population type charts showing results of tests on a number of subjects, plotted on a graph of R.sub.1 and R.sub.2, with dotted lines indicating thresholds of R.sub.1 (vertical dotted line) at 1.0 and a threshold for R.sub.2 (horizontal dotted line) at 0.66, according to one preferred embodiment of an algorithm. When both R.sub.1<1.0 and R.sub.2<0.66, the conclusion is non-concussion (lower left of chart). Also, if either the R.sub.1 or R.sub.2 is below threshold, the conclusion is non-concussion (a few in FIG. 8B have only one threshold met). If both these factors equal or exceed the threshold, i.e. the R.sub.1, R.sub.2 plot is in the upper right of the chart, concussion is indicated according to this algorithm. All three charts indicate specificity of the algorithm of the invention.

[0071] In FIG. 8A twenty-two subjects were evaluated. These were athletes in non-contact sports, aged 14 to 18. The recording sequence was two recordings every week for four weeks, for a total of 176 recordings. Of these, 159 resulted in valid recordings (due to the inability of the PPG sensor to function properly in bright sunlight). As shown on FIG. 8A, all tests come squarely in the lower left non-concussion region of the chart. Since there were zero false positives, the specificity of the test and algorithm was 100%, as to each of the 22 subjects and as to each of the 159 valid recordings. Since all subjects reported no possibly concussive events in recent months before testing or during the test, it appears probable that none of these tests represent a false negative.

[0072] FIG. 8B is another plot of R.sub.1 and R.sub.2 for 1205 recordings taken over approximately two years, from eight subjects, all of whom had no reason to believe any concussion event had occurred. No recording indicated concussion, for a specificity of 100%. As the chart shows, a few recordings were outside the lower left quadrant, with R.sub.1≥1.0 or R.sub.2>0.66, but none exceeded both thresholds. Those are considered not concussed.

[0073] FIG. 8C is a similar chart but for 1784 recordings, including the 1205 reported in FIG. 8B. The tested population included persons known to have dementia, pediatric DKA (diabetic ketone acidosis), stroke, vasospasm and 91 impact sports subjects who could possibly have had concussions, but without diagnosis. As the chart of FIG. 8C shows, 36 of the total population of 1784 recordings indicated concussion, with both R.sub.1 and R.sub.2 exceeding thresholds. Since none of the persons tested was known to have concussion, all must be assumed as non-concussed. Still, this produces a specificity of 98%, which is very high.

[0074] Note that many of the subjects reported in the testing and chart of FIG. 8C had other structural brain issues, including dementia, stroke, ischemic stroke, DKA and vasospasm. 42 had dementia and 86 had DKA, and none of these were recorded as concussed. 25 had stroke and 296 had vasospasm; only one stroke patient was recorded as concussed, and only 13 vasospasm subjects were shown as concussed. 91 of the subjects were involved in impact sports, and 12 were shown as having concussion. It is possible (but unknown) that these were true positives, but for purposes of this specificity analysis they are assumed as false positives.

[0075] FIGS. 10 and 11 show plots of accelerometer readings taken on a subject with only two opposing accelerometer sensors, at left and right temporal lobes. These are plots of acceleration events during a single heartbeat (at a heartbeat rate of 1 Hz), against time (i.e. in the time domain). FIG. 10 is from a subject without concussion; FIG. 11 is a plot from the same subject but after a concussion event. FIG. 10 indicates non-concussion, while FIG. 11 indicates concussion. The opposed positive and negative values on the charts represent data from the two opposed temporal accelerometers. Both charts have data averaged together over 45 heartbeats. From FIG. 10 it is seen that during the systole portion of a heartbeat, which is up to about 0.3 on the time scale, only two large positive peaks appear (one from the left sensor and one from the right sensor). However, the graph of FIG. 11 is strikingly different, with four prominent positive peaks showing during the systole portion of the heartbeat. This is a typical differentiating pattern to indicate non-concussion or concussion in a graph of acceleration in the time domain. Other differences occur in the remainder of the heartbeat, but the systole portion is the most prominent. As an alternative, analysis can be done by counting zero crossings with up to 8 or 10 being typical of non-concussed and a higher number being typical of concussed.

[0076] The described methodology identifies and utilizes a novel, unique pattern of cranial accelerometry that correlates with human concussion. It is influenced by movements of the brain within the skull that occur during cardiac-induced blood flow pulsations.

[0077] The terms “intensity of frequency”, “frequency content”, and “amplitude of signal data” are used in the above description and in the claims. These terms refer to how much of each of a series of particular frequencies occurs in a waveform analyzed by the algorithm, through a band of frequencies (e.g. 1-28 Hz). In the frequency domain they refer to the amplitude, i.e. height of signal data in the plot of amplitude (which can be called energy or power) versus frequency.

[0078] The above described preferred embodiments are intended to illustrate the principles of the invention, but not to limit its scope. Other embodiments and variations to these preferred embodiments will be apparent to those skilled in the art and may be made without departing from the spirit and scope of the invention as defined in the following claims.