SYSTEM AND METHOD FOR DETECTING FLUID ACCUMULATION

Abstract

A system is provided for detecting fluid accumulation in a subject. A movement induced by the beating of the heart is analyzed over time, by monitoring a movement, pressure or force. Changes in density caused by fluid accumulation result in changes in the sensor arrangement signals. Changes are used to indicate that fluid accumulation such as internal bleeding is likely to have taken place.

Claims

1. A system for detecting fluid accumulation in a body region of a subject, comprising: a sensor arrangement for obtaining a force, motion or pressure signal, induced by the beating of the heart of the subject, wherein the signal varies in dependence on the density distribution of the body region; and a processor, which is adapted to analyze changes in the sensor arrangement signals over time thereby to determine that fluid accumulation has taken place based on detecting changes in said density distribution.

2. The system as claimed in claim 1, wherein the sensor arrangement is: for location on an opposite side of the body region to the heart of the subject; or for sensing from an opposite side of the body region to the heart of the subject; or for sensing movement of an object coupled to the subject in such a way that movement of the object is induced by the beating of the heart.

3. The system as claimed in claim 1, wherein the sensor arrangement comprises: one or more pressure sensors; and/or one or more inertial sensors.

4. The system as claimed in claim 1, wherein the sensor arrangement comprises: a sensor arrangement for wearing on the head; a sensor arrangement for wearing around the waist; a sensor arrangement for wearing on the back; a sensor arrangement for wearing on the leg or feet; a sensor arrangement on which the subject is to stand, sit or lie; a non-contact sensor.

5. The system as claimed in claim 1, wherein the processor is adapted to: perform feature extraction to isolate a feature of interest of the ballistocardiography signal; determine a parameter of the feature of interest of the sensor arrangement signals; and determine from changes in the determined parameter that fluid accumulation has taken place.

6. The system as claimed in claim 5, wherein the parameter comprises a measure of signal energy, signal strength or signal variance.

7. The system as claimed in claim 5, wherein the processor comprises a decomposition algorithm for decomposing the ballistocardiography signal into frequency components, wherein the feature of interest comprises a frequency component in a frequency band which lies above a fixed lower threshold or above a determined heart rate frequency of the subject.

8. The system as claimed in claim 5, wherein: the parameter comprises a Hilbert transform of the extracted feature, and the processor comprises a Hilbert transformation unit for implementing the Hilbert transform; or the parameter comprises an energy, strength, variance or amplitude measurement of the extracted feature.

9. The system as claimed in claim 5, wherein: the parameter comprises a Hilbert transform of the raw sensor arrangement signals and the system comprises a Hilbert transformation unit for implementing the Hilbert transform; or the parameter comprises an energy, strength, variance or amplitude measurement of the raw sensor arrangement signals.

10. The system as claimed in claim 1, wherein the processor is adapted to apply a machine learning or deep learning algorithm to determine that fluid accumulation has taken place.

11. The system as claimed in claim 1, wherein the processor is adapted to analyze changes in the sensor arrangement signals over a time frame of at least two minutes.

12. The system as claimed in claim 1, wherein the sensor arrangement signals are based on: a measurement of a vector quantity or set of vector quantities in a direction parallel to the head-toe direction of the body of the subject; a measurement of a vector quantity or set of vector quantities in a direction perpendicular to the head-toe direction of the body of the subject; or a combination of multiple vector quantities in different directions.

13. The system as claimed in claim 1, wherein the signal from the sensor arrangement is filtered to only include a pre-defined human resonant frequency range falling between 2.5 Hz and 7.5 Hz.

14. A method for detecting fluid accumulation in a body region of a subject, comprising: obtaining force, motion or pressure signals, induced by the beating of the heart of the subject, wherein the signal varies in dependence on the density distribution of the body region; and analyzing changes in the signals over time thereby to determine that fluid accumulation has taken place based on detecting changes in said density distribution.

15. A non-transitory computer program comprising computer program code means which is adapted, when said program is run on a computer, to implement the method of claim 14.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0075] For a better understanding of the invention, and to show more clearly how it may be carried into effect, reference will now be made, by way of example only, to the accompanying drawings, in which:

[0076] FIG. 1 shows the relationship between an ECG signal and a BCG signal;

[0077] FIG. 2 shows a system in accordance with the invention for detecting fluid accumulation in a subject;

[0078] FIG. 3 shows raw accelerometer signals for a three-axis accelerometer over time;

[0079] FIG. 4 shows the output of the EMD process, known as IMFs;

[0080] FIG. 5 shows the Fourier Transform amplitudes for the IMFs of FIG. 4;

[0081] FIG. 6 shows an experimental setup with a pressure mat, ECG electrodes, a respiration measurement band and three-axis accelerometers;

[0082] FIG. 7 shows two different features calculated from the measurement signal;

[0083] FIG. 8 shows changes in signal characteristics using various features calculated from the short-time based analysis of the raw signal;

[0084] FIG. 9 shows the variations in energy for three different accelerometer locations;

[0085] FIG. 10 shows results for a weight measurement approach based on force sensing;

[0086] FIG. 11 shows a trend line for the IMF2 energy for the subject as a function of volume intake and the two plots normalized to the baseline;

[0087] FIG. 12 shows a method for determining that fluid accumulation has taken place; and

[0088] FIG. 13 shows the result of a frequency bases analysis of the effectiveness of the method.

DETAILED DESCRIPTION OF THE EMBODIMENTS

[0089] The invention will be described with reference to the Figures.

[0090] It should be understood that the detailed description and specific examples, while indicating exemplary embodiments of the apparatus, systems and methods, are intended for purposes of illustration only and are not intended to limit the scope of the invention. These and other features, aspects, and advantages of the apparatus, systems and methods of the present invention will become better understood from the following description, appended claims, and accompanying drawings. It should be understood that the Figures are merely schematic and are not drawn to scale. It should also be understood that the same reference numerals are used throughout the Figures to indicate the same or similar parts.

[0091] The invention provides a system for detecting fluid accumulation in a subject. A movement induced by the beating of the heart is analyzed over time, by monitoring a movement, pressure or force. Changes in density caused by fluid accumulation result in changes in the sensor arrangement signals. Changes are used to indicate that fluid accumulation such as internal bleeding is likely to have taken place.

[0092] The invention makes use of analysis of movement of the body resulting from the beating of the heart.

[0093] A ballistocardiogram (BCG) is defined as the reaction (displacement, velocity or acceleration) of the body to cardiac ejection of blood. Consequently, it is an integration of multiple forces related to movements of blood inside the heart, inside the arteries (primarily the aorta), and movement of the heart itself. Thus, this invention makes use of analysis of a signal similar to a ballistocardiogram. Whereas this type of signal is conventionally used for heart rate or respiration measurement, with the sensor placed as close to the heart as possible, the invention enables this type of motion signal to be detected remote from the heart to enable the propagation of the heart pressure wave to be detected through regions of interest of the body.

[0094] The ballistocardiogram (and the signal picked up by motion, force or pressure sensing in accordance with the invention) is inherently a multi-dimensional signal, which can be measured at all directions and all sides. These signals are strongest at the head-to-toe direction hence most BCG measurement techniques focus on the longitudinal, head-to-toe component.

[0095] It is known to use the BCG signals, in particular when implementing SCG (seismocardiography), to measure the heart of a patient and the sensing electrodes are placed as close to the heart as possible. In addition, the respiration rate can be derived from the system. The latter has the largest force in the dorso-ventral direction. However, BCG signals can be measured more generally at different sides, and at the places distant from the heart. Also they can be measured by placing the sensors on objects that the subject is touching, such as a bed, or a back support, or chair.

[0096] The invention is based on the observation that motion/pressure signal such as BCG signals have characteristics which change as a result of fluid accumulation. This happens because the movement, or vibration, or resonance of the body parts resulting from the forces generated by the beating heart and travelling blood change as a result of the fluid accumulation.

[0097] In other words, normally the internal organs vibrate in a particular manner as a result of the beating heart, but this changes because fluid is being accumulated, and their vibration characteristics change (for example damping and stiffness), which also changes the BCG signal that is measured externally to the body.

[0098] FIG. 1 shows the relationship between an ECG signal and a BCG signal as measured by a pressure mat 10 on which a subject 12 is standing.

[0099] The invention is based on the use of BCG-type information to measure internal fluid accumulation such as internal bleeding. Any suitable methods may be used to detect the movement of the body and thereby obtain the BCG-type signal. Examples are accelerometers, gyroscopes, force sensors and pressure sensors. A weighting scale for example uses load cells.

[0100] Accelerometers are already widely used in body sensors, for example in respiration sensors. Pressure sensors or load cells can be integrated in mats that can be placed in contact with a subject, for example underneath a subject. It is also known to measure BCG signals using cameras, and such systems may also be used. Cameras provide a non-contact sensing approach. Motion sensors may also be used to implement non-contact sensing, by monitoring movement of objects in contact with the subject. Such an object may be a suspended bed or a bed positioned on flexible legs, or a seat with a flexible back support. There does not need to be direct contact for example if there is a mattress between. Thus, more generally, monitoring movement of objects whose motion is caused by the BCG is possible.

[0101] In such cases, the sensor arrangement is not located on the body of the subject but is nevertheless for sensing from a known area of the body even though the sensing is contactless.

[0102] Thus, the invention is applicable for different types of sensors.

[0103] FIG. 2 shows a system in accordance with the invention for detecting fluid accumulation such as internal bleeding in a subject 12, The system comprises a sensor arrangement 20 for obtaining a sensor signal in respect of movement the subject. For convenience, the signal will from this point onwards be referred to simply as a ballistocardiography signal. The sensor arrangement 20 for example comprises an accelerometer or a pressure sensor or a combination of these, for measuring an oscillation of the body.

[0104] A processor performs analysis of the sensor arrangement signals. It may simply analyze the raw sensor arrangement signals over time for example using a machine learning or deep learning algorithm. However, to show the underlying concepts of the invention more clearly, a processing approach based on more identifiable and structured signal analysis will be described.

[0105] For this example, the processor 22 comprises:

[0106] a first signal processing unit 24 for signal pre-processing. This for example comprises filtering and decomposition enabling noise removal and extraction of the signal components;

[0107] a feature extraction unit 26 to isolate a feature of interest of the ballistocardiography signal;

[0108] a feature analysis unit 28 to determine a parameter of the feature of interest of the ballistocardiography signal. This feature of interest is in particular an energy or signal strength related feature;

[0109] an output unit 30 to determine from changes in the determined parameter that internal bleeding has taken place. This may be used to generate an alert 32 for indicating that internal bleeding has taken place.

[0110] As discussed above, organs within the body will experience oscillatory forces due to the beating of the heart, blood flow inside the large arteries (in particular the aorta) and due to breathing. To capture these oscillations, movement of the body of the patient can be measured, which will consist of movement of the patient, heart rate components and breathing components (i.e. oscillatory forces).

[0111] The beating of the heart and the movement of the torso (breathing) will induce a force that travels through the body. Any compartment (i.e. organs, tissue, fluids) that is inside the body will react differently depending on the nature of that compartment, such as the density and size, and also the damping and stiffness characteristics. When looking at the frequency domains of captured oscillatory signals, the lower frequencies originate from movement of the subject and the breathing of the patient (≈0.4 Hz), the heart rate is around 1 Hz and the frequency components above the measured heart rate are mainly the oscillations of the body due to the force exerted by the beating of the heart.

[0112] The concept is based on detecting the changes in (local) body density as a result of fluid accumulation such as internal hemorrhage. The body density can in fact change due to various reasons, such as:

[0113] accumulation of urine in the bladder;

[0114] blood leakage caused by sepsis;

[0115] consumption of food;

[0116] consumption or addition of liquids (e.g., medication or intravenous fluids);

[0117] internal hemorrhage;

[0118] inflammation.

[0119] The body density changes have different time frames and different impacts on the oscillations. For example, the consumption of food will generate a large and sudden change. This sudden change can be filtered out of the system by correcting the baseline, essentially implementing high pass filtering to remove the low frequency components.

[0120] The addition of fluids via intravenous drip is a gradual process that is redistributed across the body via the blood stream (so no accumulation of fluid) and inflammation is a relatively slow process that will change the local density of the tissue (i.e. higher perfusion, swelling). The consumption of solid food (i.e. with low water content) can also be clearly differentiated from bleeding or liquid consumption because the changes are in opposite direction.

[0121] The different processing steps used in this particular example will now be discussed in more detail.

First Signal Processing

[0122] The first signal processing unit 24 implements analysis windows of a predetermined duration, such as 5 to 60 seconds, for example 30 seconds. This predetermined duration covers a plurality of heart cycles. The windows collect sufficient data for the signal properties of interest to be measured.

[0123] These windows are analyzed for an overall analysis period, for example at least two minutes, such as 5 minutes. The raw sensor data, such as one or more accelerometer signals, in a time interval [t1 t2] is thus divided into analysis periods of 30 seconds such as [t1 t1+30], [t1+1 t1+31], and so on. The overall time period is sufficient for the fluid accumulation level to have changed.

[0124] The analysis periods may also have a partial overlap, for example windows that overlap by 50%.

Feature Extraction

[0125] The feature extraction unit 26 computes features for the signals windowed by the first signal processor. In one example, the computation comprises decomposing the original signal into its constituent frequency components. Adding these frequency components together will lead to the original signal. Since the frequency of breathing and the heart beat are significantly different, frequency decomposition leads to frequency features corresponding solely to breathing and heart beat along with other features corresponding to the oscillations due to action of the beating heart (as discussed above).

[0126] For frequency decomposition, Empirical Mode Decomposition (EMD) may be used, but alternatively wavelet transforms and filter-bank analysis can also be used. The advantage of the EMD, over these two methods is that EMD is non-parametric.

[0127] The resulting decomposition signals using EMD are called intrinsic mode functions (IMFs).

[0128] By way of example, FIG. 3 shows raw accelerometer signals for a three-axis accelerometer over time.

[0129] The raw signals are normalized to have zero mean, by dividing the signal by its mean and subtracting one, before applying EMD. The output of the EMD process are the IMFs, shown in FIG. 4.

[0130] These IMFs are time domain signals, but divided into different frequency bands, i.e. the x-axis shows time in seconds.

[0131] The number of the IMFs is dependent on the raw signal. IMFs have decreasing frequency content, the first IMF (at the top of FIG. 4) models the high frequency noise and the last IMF (at the bottom of FIG. 4) models the DC of the raw signal.

[0132] Processing of short time windows (e.g. 60 seconds) is used to calculate the IMFs. The windows are shifted in time (with a shift amount of the order of seconds, but less than the window length). Window lengths varying from 10 seconds to 60 seconds may for example be used, with comparable results. The optimal window length depends on the frequency being measured, with a longer window more suited to low frequencies and shorter windows for high frequencies.

[0133] FIG. 5 shows the Fourier Transform amplitudes for the IMFs of FIG. 4.

[0134] These Fourier Transform amplitudes are frequency domain signals, i.e. the x-axis shows frequency in Hz.

[0135] As it can be seen from FIG. 5, the first two IMFs contain high frequency components, and these are the components of interest for this application. The third (in this particular example) and later components contain the heart rate.

[0136] The frequency content of the IMFs is ordered from high frequencies to low frequencies, in other words first IMF contains the highest frequencies, and the frequency decreases as IMF number increases. The heart rate is for example visible in the third component, but that depends on the signal characteristics.

[0137] The breathing rate can be also measured. For this example, it is observed in the 6th component. Since components above the heart rate are of interest, the heart rate itself can be used to extract the right features.

[0138] As indicated above, the processing is not very sensitive to the exact choice of a cut-off frequency. This means that all components higher than 1.67 Hz or above a determined heart frequency of the subject may be used. In noisier conditions, it may be preferable to focus on selected frequency bands, and to calculate features for these bands.

Feature Analysis

[0139] The feature analysis unit 28 is used for analyzing the decomposition signals with frequencies higher than the heart rate, in order to extract representative features, such as energy. Especially the energy feature, which is linked to the signal amplitude, is associated with fluid accumulation. The energy and/or Hilbert transformation may be used to measure and study the changes in the oscillations over time.

[0140] For each one of the IMFs of FIG. 4, the energy component of the normalized time domain signal is obtained by applying an ‘Energy’ measure or ‘Hilbert’ transform.

[0141] The energy measure is the root mean squared energy of the (normalized) signal. The signal is for example normalized to have zero mean before calculating the energy. In this case (i.e. after the normalization) the energy measure is similar to the standard deviation of the signal. However, normalization of the signal before calculating the energy, is not strictly necessary.

[0142] An alternative measure, also relating to the energy content, is the Hilbert transformation applied to the chosen IMF signal.

[0143] For the calculation of the Hilbert feature, the Hilbert transform (HT) is used for local envelope detection. The HT feature is defined as the mean value of the modulus of the complex analytical signal obtained from the window under consideration. The analytical signal is defined as the sum of the complex signal, comprising the original signal and its Hilbert transform.

[0144] The Hilbert transform is essentially a transformation that makes it easy to track the envelope of a signal. The envelope is representative of variations in amplitude and thus of energy of the signal.

[0145] These features are calculated for the IMFs that have higher frequencies than the heart rate (either based on the heart rate itself, or based on a threshold level).

Output Unit

[0146] The output unit 30 implements estimation of a trend line fitting the features identified by the feature analysis. A running average is made of these features to detect any (gradual) changes in features resulting from changes inside the body. Based on the slope of the trend line, an alert may be generated (e.g., if the slope is negative an alert may be generated for potential internal bleeding).

[0147] As mentioned above, the feature analysis may focus on frequency bands higher than the resting heart rate. Experimentally, details of which are given below, it has been shown that these frequency bands are more sensitive to local changes in body density and thus can provide better distinction. An exact lower cut-off frequency is not of critical importance, as long as the frequencies lower than the heart rate are not the only ones that are included. For example, focusing on frequencies in a band is suitable with a lower threshold above the heart rate, for example of 1.67 Hz or above (for a fixed lower threshold) or based on the actual heart rate.

[0148] For the frequency domain analysis methods, such as using filter-banks, a cut-off frequency is set (e.g. in the range 1.67 Hz to 5 Hz). If the measurement of the exact heart rate is available, it can be used as an input to derive a better cut-off frequency.

[0149] The filtering is not essential. Instead, it is simply required that the higher frequencies are included in the analysis. However the detection is also possible if both higher and lower frequencies are also included, for example as is the case for raw signal processing without frequency decomposition.

[0150] Thus, there are many ways that the signal can be processed. One options is to only use frequencies higher than the heart rate and another option is to use all frequencies, also including heart rate.

Frequency Domain vs. Time Domain

[0151] The example above is based on time domain analysis, in order to derive the IMF signals of FIG. 4. The decomposition may however be performed in the frequency domain.

[0152] A time domain analysis is also possible without frequency decomposition.

[0153] For example, an energy feature may be calculated directly from the raw signal, without decomposing the raw signal into components of different frequencies. In low noise conditions, where the patient is static, this simpler method is suitable. The observed effect is present even without excluding the heart rate signal. Excluding the heart rate signal, makes the differentiation between conditions clearer, but may not be necessary.

[0154] The time domain method for example comprises the following steps:

[0155] A window period for the analysis is defined. For example, a running average window of 20 seconds may be used with calculations every second (giving a 95% window overlap). These windows are analyzed for an overall analysis period, such as 5 minutes.

[0156] For an accelerometer, one or more axes of the accelerometer may be chosen for analysis. If multiple axes are chosen, the data of the axes can be added to combine them.

[0157] The energy measure (e.g. square root of the sum of squares of each data point) is calculated, or else the Hilbert Transform of the signal may be used. Filtering or transformation of the signal may optionally be considered prior to calculation of features.

[0158] The example above is based on the decomposition of the signal into time domain signals with different frequency content and the calculation of features from one or more of the decomposed signals. These features are used to distinguish between fluid accumulation or no fluid accumulation.

[0159] The decomposition step is not always necessary. For some subjects, the difference between the conditions can be observed also without decomposing the signal into frequency bands. The method is however typically more robust with the decomposition.

[0160] Thus, in the preferred implementation, the second IMF is selected, and features are calculated from that signal. A more adaptive approach is to first determine if decomposition is needed, and if needed which frequency bands should be used. In other words, an adaptive processing method may be used, in which depending on the signal characteristics, the features are calculated from the whole signal, or from a selected set of (one or more) decomposition signals.

[0161] For both the frequency domain and time domain methods, it can be preferable to select the specific accelerometer axes, and to calculate features only for these axes. One or more of the axes can be used to reduce the noise on the signal. For example, the motion and/or external vibration of the subject can be derived from one axis and used to correct the signal along the axis capturing the head-to-toe direction. Another option is to select the channel with the highest ratio of high frequency signal energy to lower frequency signal energy.

[0162] Since the oscillations originating from the beating heart are mainly in the direction parallel to the body (i.e. from head to toes), axes in the same directions are more suitable for detecting changes. Thus, in a preferred embodiment the axis aligned with the body is selected and the energy features extracted from that signal are used.

[0163] However, in certain cases, using all accelerometer axes (i.e. summing them up) is also possible.

Experimental Results

[0164] Experiments have been performed to show that fluid accumulations in the body can be measured using this BCG based approach, in particular based on accelerometers and pressure sensors.

[0165] A water intake was used in the experiments to mimic fluid accumulation in the body. In this case, the accumulation of fluid is self-evidently in the stomach and small intestine, so the sensor arrangement is located below the stomach and small intestine, for example around the lower waist or hips. A measurement time of around 5 minutes was used, with the subject remaining still (upright or lying flat).

[0166] FIG. 6 shows the experimental setup with a pressure mat 60, ECG electrodes 62, a respiration measurement band 64 and three-axis accelerometers 66. In the workflow, a baseline was measured followed by an intake of water (of different volumes) and a subsequent measurement. The measurements were performed in both standing and lying positions.

[0167] The results based on the use of accelerometers are first discussed. They provide a more clinically applicable situation were the subject is supine. However, similar results may be obtained using weight scale measurements.

[0168] The acceleration signal in a direction parallel to the body was used.

[0169] FIG. 7 shows two different features calculated from the signal. Each feature is shown for two baseline measurements BL1,BL2 and two fluid accumulation conditions FL1 and FL2. These are signals for the same subject, in the lying position only, with repeated measurements 1 and 2.

[0170] Two different energy measuring features are calculated from the signal.

[0171] Energy is the energy measure for the whole signal in this example. i.e. without frequency decomposition.

[0172] Hilbert represents a measure of signal energy, based on calculating the Hilbert transformation of the whole signal.

[0173] Both the energy measure and the Hilbert transformation give a good distinction between the baseline (BL) and the corresponding fluid accumulation (FL) conditions. In other words, the variability of the signal monitored over time changes significantly depending on the fluid inside the body.

[0174] Due to this significant change, which is observable even for a small amount (300 ml) of accumulated fluid, changes in the signal energy or signal strength are monitored and used to generate alerts for fluid accumulation (water in the stomach and small intestine in this example, but used to validate the possibility of detecting internal bleeding).

[0175] FIG. 8 shows changes in signal characteristics (in particular the energy measure) using various features calculated from the short-time based analysis of the raw signal.

[0176] The top plot shows the mean energy feature per processing window (60 seconds) for the total measurement duration of 5 minutes. The middle plot shows the mean (the line) and standard deviation (the error bars) of the energy feature for the four test conditions (BL1, BL2, FL1, FL2), and the bottom plot shows a box-plot of the energy feature for the four test conditions. The reduction of the energy as a result of fluid accumulation is clearly visible and quantifiable from these results.

[0177] As expected, when using the heart as the source of the oscillation the position of the accelerometer has an impact on the signal.

[0178] FIG. 9 shows the variations in energy for three different accelerometer locations: acc-0 (left shoulder), acc-1 (left thigh) and acc-2 (stomach). In general, accelerometer locations between the heart and toes are particularly suitable positions based on the origin and propagation of BCG signals. The positions below the heart give the best results for this particular case, because the fluid is in the stomach. Of course for fluid accumulation above the heart, this would not be the case.

[0179] The left column is for one subject and the right column is for another subject. For each subject, BL1 and BL2 are repeated baseline measurements and FL1 and FL2 are repeated measurements for the fluid accumulation condition. They are all for the subject lying down.

[0180] FIG. 9 shows the robustness of the signal, since the measurements were conducted at different times. Note that the calibration of the accelerometers (i.e. vector analysis) and positioning will have an impact on the signal quality.

[0181] Experiments have also been performed to measure the sensitivity when using a pressure sensor while the subject is in a standing position. Measurements performed on the same subject at the same time show that signals characteristics of the pressure sensor signals are comparable to the signal characteristics of the accelerometer signals. Thus, similar sensitivity measures are applicable for accelerometer signals as for pressure signals.

[0182] FIG. 10 shows the results of signal analysis based on force measurements obtained by load cells of weight scale.

[0183] Load cells sensors in a weight scale were used to measure the BCG signal, with fluid accumulation increased by 300 ml of water. BL1 and BL2 represent two repeated baseline measurements for the subject. FL1 and FL2 are repeated measurements for the subject following a first consumption of 300 ml of water. FL3 and FL4 are a second set of two repeated measurements following a second consumption of 300 ml of water. FL5 and FL6 are a third set of repeated two measurements following a third consumption of 300 ml of water. As can be seen from FIG. 10, the weight (top left image) increases with water intake and the energy of the IMF2 (top right plot) shows a decreasing trend.

[0184] The bottom left plot shows the frequency of the IMF2 signal (x 60, i.e. in cycles per minute), and the bottom right plot shows an estimated heart rate.

[0185] The results of the two measurements show similar decrease of energy of the IMF2 signal. Even with a limited amount of repeats, the 300 ml of fluid accumulation is significantly different from the baseline.

[0186] The energy of the IMF2 signal may thus be used to create a trend line, wherein the trend is used to provide an alert that there is internal bleeding.

[0187] The top image of FIG. 11 shows, as a solid line, a trend line for the IMF2 energy for the subject represented in FIG. 10 as a function of volume intake.

[0188] The dotted line shows data for another subject (the underlying data for which is not shown). When normalizing to the baseline (volume=0) the two data sources show a good correlation with the volume intake as shown in the bottom image of FIG. 11 (which shows both plots of the top image but overlapped due to the baseline correction).

[0189] A gradient of the trend line may be used as the trigger for an alert, for example if the gradient exceeds a certain threshold (i.e. is less than a particular negative threshold value).

[0190] An amplitude difference between the baseline and the measured condition above a threshold may also be used as a trigger. In practice, this could accomplished by normalization using the baseline reading. Any deviation greater than x% (such as 2.5%) could be used to trigger an alert.

[0191] The appropriate trigger depends on the detection method. For a general feature based detection, analysis of the patterns of change or trends can be used.

[0192] For a specific feature based method, where a specific feature is correlated with the fluid accumulation in the body, a threshold based method can be used. In other words, once this feature shows particular values an alert can be generated.

[0193] For a data based model, where a machine learning approach (also including deep learning) is applied, the probability at the output of the classifier will be observed and used to make a decision to trigger an alert or not.

[0194] For a time series based approach, looking for a trend in the signal characteristics would be preferred.

[0195] FIG. 12 shows a method for detecting fluid accumulation in a subject, comprising:

[0196] in step 50, obtaining force, motion or pressure signals, induced by the beating of the heart of the subject, obtained from a location on an opposite side of the body region to the heart of the subject; and

[0197] in step 52, analyzing changes in the signals over time thereby to determine that fluid accumulation has taken place.

[0198] For the specific example above, the step 52 can be broken down to the steps of:

[0199] performing feature extraction in step 53 to isolate a feature of interest of the ballistocardiography signal;

[0200] determining in step 54 a parameter of the feature of interest of the ballistocardiography signal over a time period of multiple cardiac cycles; and

[0201] in step 56 determining from changes in the determined parameter that fluid accumulation has taken place.

[0202] An alert is then provided in step 58.

[0203] Further experiments have been performed to show that fluid accumulations in the body can be measured using this BCG based approach with better results when analyzing certain frequency ranges. In the context of BCG, it is known that the fundamental resonant frequency of the human body is in the range of 3 to 7 Hz.

[0204] The setup combined use of different sensors for acquisition of the BCG signal unobtrusively. These included:

[0205] three galvanically isolated tri-axial accelerometers;

[0206] a custom-made weighing scale consisting of four strain-gauge type pressure sensors located at four corners of a square plate, from which the total instantaneous weight of the participant was obtained by summation of the four measured values; and

[0207] an electromechanical film (EMFi) sensor mat placed on the bed and located in correspondence to the participant's torso (depending on volunteer's anatomy, approximately extending from the shoulders to the lower back).

[0208] The experiment protocol consisted of two separate test rounds, with slightly different protocols, conducted on each participant on different days, separated by up to seven days. The participants were asked to abstain from eating and drinking in the two hours preceding the test, and to empty their urinary bladder immediately before beginning of the test. In both test rounds, the initial part of the protocol was identical, and consisted in taking repeated baseline measurements in two different monitoring positions: lying on the sensor-equipped bed, and standing on the weighing scale. Each recording had a duration of 3 minutes, and standing and lying measurements were interleaved for a total of four measurements. After the initial part of the test was complete, the participant was asked to drink 500 mL of still water within a maximum 10-minute time span.

[0209] During Test Round 1, Accelerometer 1 was placed on the left side of the lower abdomen (approximately 3 cm below the navel height); Accelerometer 2 was placed on the right outer thigh (ca. 2 cm above the knee cap); Accelerometer 3 on the back of the right calf (ca. 2 cm above the ankle). During Test Round 2, Accelerometer 1 was kept at the same location as for the previous round; Accelerometer 2 was placed on the left side of the upper chest (ca. 1 cm below the collar bone); Accelerometer 3 on the right side of the lower back (around the same height as Accelerometer 1).

[0210] The experiment involved analyzing different features for the same signals, where each feature differed from other features by changing either the frequency band which was analyzed or the method of analyzing the signals. All features were initially computed on sliding windows of 10 seconds with no overlap. The complete list of features investigated includes different frequency bands and empirical mode decomposition based features.

[0211] Feature 1 (F1)—Standard deviation is obtained of the complete signal, without pre-processing.

[0212] Feature 2 (F2)—Excluding respiration band. Before computing standard deviation, the signal is filtered with a 6th-order high-pass Butterworth filter, with cutoff at 0.6 Hz. This filters the respiratory component of the BCG signal.

[0213] Feature 3 (F3)—Excluding respiration and heart-rate band. This feature is even more selective, focusing on the higher frequencies of the BCG signal. A 6th-order high-pass Butterworth with cutoff at 2 Hz is used to filter the signal before standard deviation is computed.

[0214] Feature 4 (F4)—The ‘resonance band’. This band is thought to contain those frequencies of the BCG signal which closely match the body's natural frequencies, thus entering a condition of resonance which enhances the response vibration. The pass-band for this feature has been defined to have cutoffs at 2.5 and 7.5 Hz. The filter used is again a 6-th order Butterworth.

[0215] Feature 5 (F5)—Standard deviation denoised with Empirical Mode Decomposition (EMD).

[0216] Feature 6 to 9 (F6-F9)—Standard deviations of, respectively, IMF 2, 3, 4 and 5. These features are based on EMD and selectively compute the standard deviations.

[0217] Feature 10 to 18 (F10-F18)—median absolute deviation (MAD) based feature set. These nine features involve the same pre-processing steps as features 1 to 9, but compute the energy using MAD instead of the standard deviation. The MAD is defined as the median of the absolute deviations from the data's median. It is a measure similar to standard deviation, however it is more robust for outliers than standard deviation.

[0218] All features were calculated from each of the 12 signals available (the 8 accelerometer channels, EMFi and the additional 3 signals obtained by the sum of the axes for each accelerometer).

[0219] FIG. 13 shows the decrease for each feature from the baseline measurement to the measurement after ingestion according to a t-test (one type of hypothesis test) calculated from the average among the 12 signals. It appears that the most promising features are Feature 4 (i.e. resonance based feature) and the respective MAD based Feature 13.

[0220] Thus, one implementation of the system may involve 13 filtering the signal from the sensor arrangement is filtered to only include a pre-defined human resonant frequency range falling between 2.5 and 7.5 Hz, for example 3 Hz to 7.5 Hz, for example 2.5 Hz to 7.0 Hz, for example 3 Hz to 7 Hz.

[0221] Although the main application is internal bleeding, other application areas are foreseen. One of those application is the fluid accumulation/redistribution that is not related to bleeds. For example the fluid accumulation in heart failure patients, or the redistribution of fluids during pregnancy. Here, also, the measurement in a standing position is feasible. For a measurement in standing position, a feature band with a lower value of 5 Hz and an upper value of 11 Hz and/or 12 Hz and/or 13 Hz may be preferred. This is a slight shift with respect to the lying down position. This shift is due to the mechanical response of the body that change based on the compression of the tissue/organs/sections on the body. One can see the body as a set of elements linked by ‘springs’ and ‘dampers’ which are not linear during compression.

[0222] In order to determine the optimal band, a frequency sweep around the start and end frequencies of the feature 4 may be performed, and the best performing frequency band selected as the final feature. To have a robust feature applicable for different body positions and participants, the widest possible frequency band, without degrading the performance, can be selected.

[0223] Based on a combination of a BCG and a Lumped-Mass Mechanical model of a human, a personalized resonance frequency can be calculated based on the weight, body shape (partly also incorporated in male/female difference) and length. This model gives a generalized relation for the resonance frequency as function of body mass and body mass index. The gender and age could also be incorporated into the model above..

[0224] Adaptation of the resonance frequency for the positions can be also performed. (Such adaptation is possible, by collecting representative experimental data and learning from it, or by using existing (different body position) models, simulating them with experimental data, and analyzing the simulation results.)

[0225] It is known that humans in different body positions have slightly different resonant frequencies. In order to increase the accuracy and detection sensitivity, the position of the subjects (e.g. supine, standing, sitting, etc.) can be estimated using accelerometers, then using the position information, a personalized resonance band can be determined. Such adaptations can make the detection algorithm more specific and targeted for specific conditions, which may enable higher accuracy and sensitivity.

[0226] As discussed above, the system makes use of a processor to perform the data processing. The processor can be implemented in numerous ways, with software and/or hardware, to perform the various functions required. The processor typically employs one or more microprocessors that may be programmed using software (e.g., microcode) to perform the required functions. The processor may be implemented as a combination of dedicated hardware to perform some functions and one or more programmed microprocessors and associated circuitry to perform other functions.

[0227] Examples of circuitry that may be employed in various embodiments of the present disclosure include, but are not limited to, conventional microprocessors, application specific integrated circuits (ASICs), and field-programmable gate arrays (FPGAs).

[0228] In various implementations, the processor may be associated with one or more storage media such as volatile and non-volatile computer memory such as RAM, PROM, EPROM, and EEPROM. The storage media may be encoded with one or more programs that, when executed on one or more processors and/or controllers, perform the required functions. Various storage media may be fixed within a processor or controller or may be transportable, such that the one or more programs stored thereon can be loaded into a processor.

[0229] A time-series based processing method is discussed above, but it is also mentioned that machine learning may be used. A machine learning classifier would continuously generate a probabilistic score as to whether fluid accumulation is detected or not. Deep learning based methods may also be used.

[0230] Variations to the disclosed embodiments can be understood and effected by those skilled in the art in practicing the claimed invention, from a study of the drawings, the disclosure and the appended claims. In the claims, the word “comprising” does not exclude other elements or steps, and the indefinite article “a” or “an” does not exclude a plurality. A single processor or other unit may fulfill the functions of several items recited in the claims. The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage. A computer program may be stored/distributed on a suitable medium, such as an optical storage medium or a solid-state medium supplied together with or as part of other hardware, but may also be distributed in other forms, such as via the Internet or other wired or wireless telecommunication systems. Any reference signs in the claims should not be construed as limiting the scope.