METHOD FOR DETERMINING ACCURACY OF HEART RATE VARIABILITY

Abstract

A computer implemented method for determining accuracy of heart rate variability is proposed. The method comprises the following steps: a) providing at least one photoplethysmogram obtained by at least one portable photoplethysmogram device (110); b) Determining at least one signal feature by evaluating the photoplethysmogram; c) Determining the accuracy of heart rate variability by using at least one trained model, wherein the signal features determined in step b) are used as input for the trained model.

Claims

1. A computer implemented method for determining accuracy of heart rate variability comprising the following steps: a) providing at least one photoplethysmogram obtained by at least one portable photoplethysmogram device; b) determining at least one signal feature by evaluating the photoplethysmogram; c) determining the accuracy of heart rate variability by using at least one trained model, wherein the signal features determined in step b) are used as input for the trained model; wherein the accuracy is used for distinguishing between acceptable and non-acceptable heart rate variability data, wherein the method comprises comparing the accuracy to at least one threshold, wherein, if the accuracy is below the threshold, a heart rate variability data point is considered as acceptable, otherwise as non-acceptable.

2. The method according to claim 1, wherein the accuracy is used as quality indicator for heart rate variability data.

3.-4. (canceled)

5. The method according to claim 1, wherein the method comprises determining the at least one threshold.

6. The method according to claim 1, wherein the photoplethysmogram comprises at least one signal, wherein the method comprises evaluating the signal, wherein the evaluation comprises one or more of interpolating the signal, resampling the signal, isolating signal component, analyzing considering non-overlapping windows, normalizing, identifying of peaks.

7. The method according to claim 1, wherein the signal feature comprises at least one feature selected from the group consisting of: root mean square of successive differences (RMSSD), standard deviation of the R-to-R intervals (RRI) (SDNN), standard deviation of the RRIs in a current window, pnn50 from the photoplethysmogram (PPG), average heart rate from PPG in the current window, number of ectopic RRIs in the current window, minimum RRI value in the current window, variance of the RRIs in the current window, number of RRIs in the current window, 95.sup.th percentile of the RRIs in the current window, 5th percentile of the RRIs in the current window, variance of a raw PPG signal in the current window, max value of the raw PPG signal in the current window, min value of the raw PPG signal in the current window, average value of the raw PPG signal in the current window, standard deviation of the raw PPG signal in the current window, entropy of the raw PPG signal in the current window, kurtosis of the raw PPG signal in the current window, skewness of the raw PPG signal in the current window, variance of the filtered PPG signal in the current window, max value of the filtered PPG signal in the current window, min value of the filtered PPG signal in the current window, average value of the filtered PPG signal in the current window, standard deviation of the filtered PPG signal in the current window, kurtosis of the filtered PPG signal in the current window, skewness of the filtered PPG signal in the current window.

8. The method according to claim 1, wherein the trained model comprises at least one model selected from the group consisting of: a linear regression model, e.g. comprising transformed features, such as log-transformed or polynomial; at least one non-linear Artificial Neural Network (ANN); at least one Support Vector Machine (SVM); at least one kernel based method; Tree regression; Random Forest.

9. The method according to claim 1, wherein the method comprises at least one training step, wherein, in the training step, the trained model is trained on at least one training dataset, wherein the training dataset comprises a set of heart rate variability values determined by using at least one electrocardiogram device and heart rate variability values determined by using the photoplethysmogram device.

10. The method according to claim 9, wherein the method comprises determining at least one heart rate variability metric of the heart rate variability values determined by using at least one electrocardiogram device and determining at least one heart rate variability metric of the heart rate variability values determined by using the photoplethysmogram device, wherein the method comprises comparing the heart rate variability metrics against each other.

11. The method according to claim 10, wherein the method comprises determining at least one error of heart rate variability by combining the heart rate variability metric determined by using at least one electrocardiogram device and the heart rate variability metric of the heart rate variability values determined by using the photoplethysmogram device, wherein the error of heart rate variability is used together with signal features extracted from the photoplethysmogram for determining the trained model for determining the heart rate variability accuracy.

12. The method according to claim 1, wherein the photoplethysmogram device comprises at least one illumination source and at least one photodetector.

13. A portable photoplethysmogram device, wherein the portable photoplethysmogram device is configured for determining accuracy of heart rate variability, wherein the portable photoplethysmogram device comprises at least one illumination source and at least one photodetector configured for determining at least one photoplethysmogram, the portable photoplethysmogram device further comprises at least one processing unit configured for determining at least one signal feature by evaluating the photoplethysmogram, wherein the processing unit is configured for determining the accuracy of heart rate variability by using at least one trained model, wherein the determined signal features are used as input for the trained model, wherein the accuracy is used for distinguishing between acceptable and non-acceptable heart rate variability data, wherein the portable photoplethysmogram device is configured for comparing the accuracy to at least one threshold, wherein, if the accuracy is below the threshold, a heart rate variability data point is considered as acceptable, otherwise as non-acceptable.

14. The portable photoplethysmogram device according to claim 13, wherein the portable photoplethysmogram device is configured for performing the method according to claim 1.

15. A computer program comprising instructions which, when the program is executed by the portable photoplethysmogram device according to claim 13 referring to a portable photoplethysmogram device, cause the portable photoplethysmogram device to carry out steps a) to c) of the method according to claim 1.

16. A computer-readable storage medium comprising instructions which, when executed by the portable photoplethysmogram device according to claim 13 referring to a portable photoplethysmogram device, cause the portable photoplethysmogram device to carry out steps a) to c) of the method according to claim 1.

Description

IN THE FIGURES

[0097] FIG. 1 shows a flow diagram of the method and at least one portable photoplethysmogram device according to the present invention;

[0098] FIGS. 2A and 2B show an example of raw Electrocardiogram data and evaluated Electrocardiogram data;

[0099] FIGS. 3A and 3B show an example of raw photoplethysmogram data and evaluated photoplethysmogram data;

[0100] FIGS. 4A to 4C show an example of determining of R-to-R intervals;

[0101] FIGS. 5A to 5F show heart rate variability accuracy results obtained for different threshold levels; and

[0102] FIGS. 6A to 6D show in FIG. 6A to 6C True Positive Rate (TPR) and False Positive Rate (FPR) when binary classifying heart rate variability values depending on possible threshold levels and in FIG. 6D RMSE (left) and MARD (right) accuracy metrics for SDNN values considered as False Positives (FP) when a given value of the threshold is used on the estimated multivariate heart rate variability error metric to distinguish between accurate and inaccurate readings.

DETAILED DESCRIPTION OF THE EMBODIMENTS

[0103] FIG. 1 shows a flow diagram of the method for determining accuracy of heart rate variability and at least one portable photoplethysmogram device 110 according to the present invention. The heart rate variability (HRV) may be a measure of regularity between consecutive heartbeats.

[0104] The photoplethysmogram device 110 is configured for determining at least one photoplethysmogram (PPG). The PPG may show development of a signal from the PPG device 110 over time.

[0105] The photoplethysmogram device 110 comprises at least one illumination source 112. The illumination source 112 may comprise at least one light source such as at least one light-emitting-diode (LED) transmitter. The illumination source 112 may be configured for generating at least one light beam for illuminating e.g. the skin on at least one part of the human body. The illumination source 112 may be configured for generating light in the red, infrared or green spectral region.

[0106] The photoplethysmogram device 110 may comprise at least one photodetector 114, in particular at least one photosensitive diode. The photodetector 114 may be configured for detecting a light beam, such as for detecting an illumination generated by at least one light beam. The photodetector 114 may be configured for detecting light from transmissive absorption and/or reflection in response to illumination by the light generated by the illumination source 112.

[0107] The PPG may comprise a plurality of beats. The heart rate variability may be measured by the variation in the beat-to-beat intervals, also denoted R-to-R intervals (RRI). Generally, an R wave is a section of an Electrocardiogram (ECG) signal consisting of a sharp raise followed by a sharp decrease of the signal. The morphology of a PPG signal may be different from the ECG one but still showing repetitive pattern due to heart beats. The heart rate variability may be defined as the variation of successive heartbeats.

[0108] The accuracy may be a measure for closeness of a measurement value to a certain value, in particular a true value. The true value may be a heart rate variability value determined using at least one ECG device 116.

[0109] The PPG device 110 may be wearable. For example, the PPG device 110 may be a wristwatch such as a smartwatch. Using a portable PPG device 110 may result in that disturbances can influence the HRV measurement such as motions artefacts. Uncontrolled conditions met in daily life may pose several challenges related to disturbances that can deteriorate a PPG signal 118 making the calculation of the HRV untrustworthy and not reliable.

[0110] The signal 118 may be at least one electronic signal of the PPG device 110, in particular of the photodetector 114, depending on detected light from transmissive absorption and/or reflection in response to illumination by the light generated by the illumination source 112.

[0111] The PPG device 110 may further comprise at least one processing unit 120 configured for determining at least one signal feature by evaluating the photoplethysmogram. The step of feature extraction is denoted with reference number 121. The photoplethysmogram may comprises at least one signal, also denoted as PPG signal 118. The evaluation of the PPG signal 118 may comprise one or more of interpolating the signal, resampling the signal, isolating signal components, analyzing considering non-overlapping windows, normalizing, identifying of peaks.

[0112] For example, the PPG signal 118 may be interpolated over a uniform time grid to account for slight fluctuations of sampling frequency, such as around 20 Hz. The PPG signal 118 may be resampled to increase the sampling frequency, such as to 1 kHz, for example by using an averaging filter of length 0.5 seconds and a Blackman window.

[0113] The PPG signal 118 may comprise a slow trend, often referred to DC component. Without being bound by this theory, this trend is likely due to respiration and other low frequency physiology-related modulations, see Julien, Claude. “The enigma of Mayer waves: facts and models.” Cardiovascular research 70.1 (2006): 12-21. The PPG signal 118 may comprise a pulsatile component, often referred to AC, due to blood volume variations synchronized with the heart beats. FIG. 4A shows a further example of a raw PPG signal 118 under rest conditions where the components are visible. To isolate the AC component from the PPG signal 118 a Morlet wavelet may be used, i.e.

[0114] a very selective band pass filter, centered around the frequency of interest, i.e. heart rate. For additional details about the Morlet wavelet reference is made toCohen, Michael X. “A better way to define and describe Morlet wavelets for time-frequency analysis.” NeuroImage 199 (2019): 81-86. It was found that the accuracy values are the key to make sure that the filtered signal contains the pulsatile component and not, for example, motion artefacts.

[0115] The PPG signal 118, in particular the resampled and interpolated PPG signal, may be analysed considering non-overlapping windows, such as windows of 20 seconds. For each window a median heart rate may be determined. The method may comprise using the median heart rate to build wavelet filter coefficients. Before applying the filter, a PPG waveform in a current window may be normalized by a DC mean value.

[0116] The peaks on the filtered PPG signal 118 may then be identified and/or determined and/or calculated looking at a combination of first and second derivatives of the signal. Identified peaks may then be concatenated until the last window that has been analyzed. A RRI time series, i.e. a specific number of consecutive peaks, may be filtered with a heuristic rule to make sure erroneous beats are excluded from the calculation of the HRV statistics. For example, a current RRI may be kept when it differs less than 30% from the previous one and the previous one, i.e. differs less than 30% from the one before, i.e. RII.sub.i-2. Otherwise the RRI may be removed from the RRI time series.

[0117] The signal features may comprise both statistics describing the PPG signal 118 as well as statistics describing the RRI distributions. The former ones may comprise one or more of variance, minimum, maximum, average, standard deviation, entropy, kurtosis and skewness values over raw and filtered PPG signals 118. The latter ones may comprise one or more of average RRIs and HR, the absolute number of filtered RRIs and the ratio between good and filtered RRIs, the minimum and maximum number of RIIs as well as the 5th and 95th RRI percentiles. The signal feature may be determined for a current time instant t.sub.i considering RRIs temporally located between the current time instant t.sub.i and a time instant t.sub.i-wl, wherein wl is a window length ranging from seconds, e.g. 30 seconds, to minutes, such as 5 minutes. The signal feature comprise at least one feature selected from the group consisting of: root mean square of successive differences (RMSSD), standard deviation of the RRI intervals (SDNN), standard deviation of the RRIs in a current window, pnn50 from PPG, root mean square of pnn50, average RRI value from PPG in the current window, average heart rate from PPG in the current window, number of ectopic RRIs in the current window, ratio between number of ectopic and normal RRIs in the current window, minimum RRI value in the current window, maximum RRI value in the current window, variance of the RRIs in the current window, number of RRIs in the current window, 95th percentile of the RRIs in the current window, 5th percentile of the RRIs in the current window, variance of a raw, i.e. not filtered, PPG signal 118 in the current window, max value of the raw PPG signal 118 in the current window, min value of the raw PPG signal 118 in the current window, average value of the raw PPG signal 118 in the current window, standard deviation of the raw PPG signal in the current window, entropy of the raw PPG signal 118 in the current window, kurtosis of the raw PPG signal 118 in the current window, skewness of the raw PPG signal 118 in the current window, variance of the filtered PPG signal 118 in the current window, max value of the filtered PPG signal in the current window, min value of the filtered PPG signal in the current window, average value of the filtered PPG signal in the current window, standard deviation of the filtered PPG signal 118 in the current window, kurtosis of the filtered PPG signal 118 in the current window, skewness of the filtered PPG signal 118 in the current window. In the method according to the present invention all of these signal features may be determined or a subset of these signal features may be determined. It was found that the following subset of features is particular suitable for a reliable determination of accuracy of heart rate variability: root mean square of successive differences (RMSSD), standard deviation of the RRI intervals (SDNN), standard deviation of the RRIs in a current window, pnn50 from PPG, average heart rate from PPG in the current window, number of ectopic RRIs in the current window, minimum RRI value in the current window, variance of the RRIs in the current window, number of RRIs in the current window, 95.sup.th percentile of the RRIs in the current window. The RMSSD may be determined by calculating the square root of the mean of the squares of the successive differences of consecutive RRIs:

[00008]$\mathrm{RMSSD}=\sqrt{\frac{{{\Sigma }_{i=0}^{N-1}\left({\mathrm{RRI}}_i-{\mathrm{RRI}}_{i+1}\right)}^2}{N-1}}.$

[0118] The SDNN may be determined by calculating:

[00009]$\mathrm{SDNN}=\sqrt{\frac{{{\Sigma }_{i=1}^N\left({\mathrm{RRI}}_i-\stackrel{\_}{\mathrm{RRI}}\right)}^2}{N-1}},$

[0119] where is RRI the average of the RRI in the considered time window. pnn50 is the proportion of NN50 divided by total number of RRIs, wherein NN50 is the number of pairs of successive RRIs that differ by more than 50 ms.

[0120] The processing unit 120 is configured for determining the accuracy of heart rate variability by using at least one trained model, wherein the determined signal features determined are used as input for the trained model. The steps shown inside box 122 of FIG. 1 may be performed by the processing unit 120.

[0121] The method may comprise at least one training step, wherein, in the training step, the trained model is trained on the at least one training dataset. The steps outside and inside the box 122 may be performed during the training step. The trained model may comprise at least one model selected from the group consisting of: a linear regression model, e.g. comprising transformed features, such as log-transformed or polynomial; at least one non-linear Artificial Neural Network (ANN), in particular at least one deep learning architecture such as Convolutional NN, Recurrent NN, Long Short Term Memory NN, and the like; at least one Support Vector Machine (SVM); at least one kernel based method; Tree regression; Random Forest.

[0122] The training dataset may comprise of a set of HRV values determined by using the ECG device 116 and HRV values determined by using the PPG device 110. ECG and PPG data may be collected simultaneously.

[0123] For example, for the experimental results shown in FIGS. 2 to 6, the training dataset consists of 20 recordings where ECG and PPG data are collected simultaneously from 20 healthy volunteers (4 female and 16 male) while performing a series of activities. The training dataset may be determined by performing at least one test protocol comprising the series of activities. During the test protocol, subjects were wearing a 3-LEDs ECG device 116 with sampling frequency at 1 kHz (BioRadio, by Neurotechnologies) and as PPG device 110 a smart watch on the wrist equipped with LEDs and photodiode for measuring PPG at 20 Hz (Samsung™ Gear Sport Smartwatch).

[0124] The protocol may comprise of a series of activities meant to induce HRV variations so to compare HRV over a wide range of values as well as inducing motion artefacts to test the ability of the algorithm and the quality metric to distinguish between accuracy and inaccurate HRV values. Some protocol activities, e.g. console gaming, mental stress manipulation and physical activity, may be included to reflect typical activities performed in daily life use of the PPG device. Pace breathing may be considered because it increases the range of HRV values through respiratory sinus arrhythmia, allowing the calculation of results over a broad range of variation and making easier the post alignment/synchronization of the time series obtained from the reference ECG and the PPG signals. The following table gives a list of an exemplary protocol:

TABLE-US-00002 Activity Duration Screening & Informed consent process (while sitting, — at rest) Placement of ECG and PPG sensors (while sitting, — at rest) Baseline (sitting, at rest) 5 minutes Paced breathing (ladder of increasing respiratory 5 minutes frequencies from 5 to 20 breaths per minute with steps of 5) Console gameplay (PS4 Aaero) 5 minutes Orthostasis (standing, otherwise at rest) 5 minutes Mental stress manipulation (Serial 7s [subtraction 5 minutes by 7 from 700, with eyes closed, pronouncing aloud each response]) Physical activity manipulation (uninterrupted indoor 5 minutes walking along a pre-set circular path; same path for all subjects) Baseline (sitting, at rest) 5 minutes Retrieve PPG/ECG equipment and debrief —

[0125] The method may comprise analyzing ECG and PPG data to obtain the RRIs time series from which heart rate variability metrics can be derived. The method may comprise comparing the heart rate variability metrics against each other to obtain a measure of the accuracy.

[0126] The PPG signals may be evaluated as described above. The evaluation step is denoted with reference number 124. The signal features from the PPG signal and from the ECG signal may be calculated over the same time window. FIG. 2A shows an example of raw PPG data and evaluated, in FIG. 2B, with the proposed wavelet based algorithm to improve heart beats detection.

[0127] For the ECG data 126 comprising a plurality of ECG signals, a different evaluation may be performed. The raw ECG signal may be analyzed with a variation on the Pan-Tompkins algorithms, see Pan J, Tompkins W J. A real-time QRS detection algorithm. IEEE Trans Biomed Eng. 1985 March; 32(3):2. A Savitzky-Golay differentiation filter may be used to provide a filtered version of the raw signal first derivative, see Savitzky, A., Golay, M. J. E. “Smoothing and Differentiation of Data by Simplified Least Squares Procedures”. Analytical Chemistry. 1964, 36(8): 1627-39.doi:10.1021/ac60214a047. The ECG signal may be squared for emphasizing higher frequencies and filtered with a moving integrator filter, e.g. of width 60 ms, i.e. the average QRS complex width, for obtaining the ECG shape back with highlighted QRS complexes. The signal may be normalized with its envelope that is obtained at each time instant by filtering the root mean square of the signal in a rolling window of length Fs/2 with a Butterworth low pass filter with cutoff frequency at, e.g. 0.8 Hz, where Fs represents the sampling frequency of the ECG signal. Single heart beats may be identified on this normalized signal as the peaks exceeding a threshold that in our case was identified as the 90th percentile of the data in the current window. Each heart beat crossing the threshold may be subsequently checked manually to make sure no erroneous beat was included in the analysis. The evaluation of the ECG data 126 is shown with reference number 128 in FIG. 1. FIG. 3A shows an example of raw ECG data and evaluated, FIG. 3B, with the proposed algorithm to improve heart beats detection.

[0128] FIG. 4B shows a further example, of raw and filtered ECG signal with the Pan-Tompkins algorithm. FIG. 2A and FIG. 2B are combined onto FIG. 4B. The Figures refers to different time windows. In addition, in FIG. 4B the location of the peaks are not shown. Same for FIGS. 3A and 3B, that are combined into FIG. 4A. FIG. 4C shows a comparison of respective RRI intervals for a representative 2 minutes window for the PPG of FIG. 4A and the ECG of FIG. 4B.

[0129] The method may comprise determining at least one heart rate variability metric of the heart rate variability values determined by using at least one electrocardiogram device 116, denoted with reference number 130 in FIG. 1. The method may comprise determining at least one heart rate variability metric of the heart rate variability values determined by using the PPG device 110, denoted with reference number 132. The heart rate variability metric may be statistics calculated on RRIs contained inside a time window of specific length that can last from minutes to hours. For further details about the heart rate variability metric reference is made to Fei, Lu, et al, “Short- and long-term assessment of heart rate variability for risk stratification after acute myocardial infarction.” The American journal of cardiology 77.9 (1996): 681-684 and Mourot, Laurent, et al. “Short- and long-term effects of a single bout of exercise on heart rate variability: comparison between constant and interval training exercises.” European journal of applied physiology 92.4-5 (2004): 508-517. In particular, the heart rate variability metrics may be calculated on time window of minutes, what in the literature is sometimes referred as “short-term” heart rate variability. The heart rate variability metrics may be calculated in the specific 30, 60, 90, 120, 180, 240 and 300 seconds. Heart rate variability metrics may belong to different classes depending on the domain of the method used to analyze the RRIs.

[0130] For example, the heart rate variability metrics may comprise the time, frequency, non-linear and geometrical domains. The heart rate variability metrics may comprise the Root Mean of the Squared Differences (RMSSD) of consecutive RRIs:

[00010]$\mathrm{RMSSD}=\sqrt{\frac{{{\Sigma }_{i=0}^{N-1}\left({\mathrm{RRI}}_i-{\mathrm{RRI}}_{i+1}\right)}^2}{N-1}}$

[0131] and the Standard Deviation of the NN intervals (SDNN):

[00011]$\mathrm{SDNN}=\sqrt{\frac{{{\Sigma }_{i=1}^N\left({\mathrm{RRI}}_i-\stackrel{\_}{\mathrm{RRI}}\right)}^2}{N-1}}$

[0132] where is RRI the average of the RRI in the considered time window. Another metric derived from the interval differences may comprise the PNN50, that is, the number of consecutive RRIs differing more than 50 ms normalized by the total number of RRIs in the considered window.

[0133] The heart rate variability metrics obtained from the PPG and the ECG may be combined to define a heart rate variability error, denoted with reference number 134. The method may comprise determining at least one error of heart rate variability, i.e. the difference between heart rate variability values obtained with the PPG and the ECG. Specifically, for each heart rate variability metric an error may be determined, at each time instant i-th, as the absolute difference between the heart rate variability values obtained from the PPG and ECG signals:

Err.sub.SDNN.i=|SDNN.sub.ECG,i−SDNN.sub.PPG,i|.

[0134] The method may comprise considering a combination of heart rate variability error metrics where at each time instant i-th, the multivariate error metric Err.sub.multivariate,i is the average of the errors Err.sub.SDNN.i for each heart rate variability metric.

[0135] The error of heart rate variability (HRVE) may be used together with signal features extracted from the PPG for determining the trained model for determining the heart rate variability accuracy itself, denoted with reference number 136. The method may comprise performing at least one multivariate supervised regression, wherein as input the at least one signal feature extracted from the photoplethysmogram may be used. The output may be the error between the heart rate variability metrics obtained from the PPG signal 118 and the ones obtained from the ECG signal.

[0136] For example, as model a linear model in the form may be used:

HRVE=Xβ

[0137] where HRVE is a (n×1) vector collecting the HRVE.sub.i values Err.sub.multivariate.i, X is the (n×p) matrix collecting the features obtained from the PPG and β is the (p×1) vector collecting the model coefficients. The i-th row of matrix X collects the p features calculated in the same time window of PPG data that is used to calculate the i-th heart rate variability value. For example, as estimation technique a Least Absolute Shrinkage and Selection Operator (LASSO) may be used. These techniques may comprise a L1 norm regularization and has the property of setting to zero coefficients in the model associated with unimportant features, allowing to control for complexity and avoiding overfitting, see Tibshirani R. Regression Shrinkage and Selection via the lasso. Journal of the Royal Statistical Society. Series B (methodological). 1996 58(1): 267-88.

[0138] In one embodiment the HRV accuracy may be trained with a subset of signal features:

[00012]$HR{V}_{\mathrm{ACCURACY}}\left(t_i\right)={\beta }_{1}rmss{d}_{ppg}\left(t_i\right)+{\beta }_{2}pnn50_{ppg}\left(t_i\right)+{\beta }_3{\mathrm{avg\_hr}}_{PPG}\left(t_i\right)+{\beta }_4\mathrm{n\_ectpc}{\mathrm{\_rri}}_{ppg}\left(t_i\right)+{\beta }_5\mathrm{min\_rri}\mathrm{\_ppg}\left(t_i\right)++{\beta }_6{\mathrm{var}}_{{\mathrm{rri}}_{ppg}\left(t_i\right)}+{\beta }_7{\mathrm{std}}_{rr{i}_{ppg}\left(t_i\right)}+{\beta }_8{n}_{{\mathrm{rri}}_{ppg}\left(t_i\right)}+{\beta }_{9}95{\mathrm{perc}}_{rr{i}_{ppg}\left(t_i\right)},$

[0139] wherein β.sub.j are the respective model coefficients, rmssd_ppg is the RMSSD from the PPG, pnn50.sub.ppg is the pnn50 from the PPG, avg_hrp.sub.PPG is the average heart rate from the PPG in the current window, n_ectpc_rri.sub.ppg is the number of ectopic RRIs in the current window, min_rri_ppg is the minimum RRI value in the current window, var_rri_ppg is the variance of the RRIs in the current window, std_rri_ppg is the standard deviation of the RRIs in the current window, n_rri_ppg is the number of RRIs in the current window and 95perc_rri_ppg is the 95th percentile of the RRIs in the current window.

[0140] The method, in particular the training step, may comprise at least one validation step, wherein a Leave-One-Subject-Out Cross-Validation (LOSO-CV) is used. At each iteration N−1 subjects out of N subjects may be used to train the model. In the validation step, trained model is tested on the data from the subject that was left out from the training dataset, see Friedman, Jerome, Trevor Hastie, and Robert Tibshirani, The elements of statistical learning. Vol. 1. No. 10. New York: Springer series in statistics, 2001.

[0141] The trained model identified in step 136 can be used to estimate the heart rate variability error, e.g. prospectively, when ECG data is not available. This step is denoted with reference number 140 in FIG. 1.

[0142] The determined accuracy may be used as quality indicator for heart rate variability data. A better accuracy should be associated with a high quality and a low accuracy with a bad quality. The accuracy may be reflected by a quality metric. The heart rate variability determined from the photoplethysmogram may be an actual value in the quality metric. The quality metric may set a tolerance range that defines acceptable data points. The quality metric may be used for deciding and/or differentiating and/or distinguish between acceptable and non-acceptable heart rate variability data points.

[0143] The method may comprise comparing the accuracy to at least one threshold. If the accuracy is below the threshold, a heart rate variability data point may be considered as acceptable, otherwise as non-acceptable. The threshold may be used to distinguish between acceptable and unacceptable heart rate variability values. The method may comprise a binary decision to include or exclude a heart rate variability data point, denoted with reference number 142 in FIG. 1. The result of this decision is denoted as “HRVQ” in FIG. 1. The threshold may be a pre-determined or pre-defined threshold. The threshold may be set on the continuous HRVE values estimated by the trained model. When HRVE is below the threshold the respective heart rate variability value or data point may be considered with good quality and as “acceptable”, otherwise it is not and is considered as “non-acceptable”. An acceptable heart rate variability value may have a heart rate variability quality equal to 1 and a non-acceptable heart rate variability value may have a heart rate variability quality equal to 0.

[0144] The method may comprise determining the threshold, in particular at least one threshold level. Influences of different threshold levels may be tested as follows. For example, for all the considered heart rate variability metrics, the calculation of the heart rate variability accuracy may be performed using at least one performance metrics as a function of the threshold levels. Additionally or alternatively, the influences may be tested using an analysis considering errors arising from setting a threshold on a continuous value, HRVE, which is estimated by a model and thus presents uncertainty. The analysis may thus be highly dependent on the ability of the trained model to accurately predict HRVE. For example, a Receiver Operating Characteristic (ROC) analysis may be used using the true and the predicted values of HRVE for different threshold levels. For each threshold value, a confusion matrix may be calculated, a True Positive Rate (TPR), i.e. the rate of good quality HRV values classified as such, may be determined and a False Positive Rate (FPR), i.e. the number of inaccurate HRV values that are nevertheless included in the analysis because of the uncertainty in the predicted HRVE, may be determined. The heart rate variability accuracy of those points identified as FPR may have an indication of heart rate variability accuracy degradation derived from including these points.

[0145] FIGS. 5A to 5E show the HRV accuracy results for the multivariate error metric when the window length is 120 seconds obtained for different threshold levels, including the accuracy of pulse rate and the percentage of good data (relative to the total amount of data) included in the analysis as a function of the threshold. FIG. 5E shows the mean absolute error “MAE”. In general, the higher the threshold the more data is considered as accurate (see FIGS. 5F), but the accuracy of the HRV metrics decreases.

[0146] The vertical line 144, at threshold value around 30, is the accuracy used by FDA to clear a device for pulse rate monitoring as a medical device, see ANSI/AAMI EC13-1992, “Cardiac monitors, heart rate meters, and alarms”. This threshold gives an error in terms of RMSE for RMSSD around 30 ms and for SDNN around 15 ms. A threshold at 20 may be more desirable since the RMSE would drop below 15 ms for SDNN and around 20 for RMSSD.

[0147] Errors in the prediction of the HRV quality could cause the inclusions of data points that are actually inaccurate, as well as the exclusion of points that are accurate. To test what is the influence of these type of errors on the overall HRV accuracy the predicted and real values of HRVE were used to build the ROC curves in FIG. 6. FIGS. 6A to 6C show True Positive Rate (TPR) and False Positive Rate (FPR) when binary classifying HRV values as accurate or inaccurate depending on possible threshold levels and FIG. 6D shows RMSE (left) and MARD (right) accuracy metrics for SDNN values considered as False Positives (FP) when a given value of the threshold is used on the estimated multivariate HRV error metric (HRVQ) to distinguish between acceptable and non-acceptable readings. FIG. 6B shows that with a threshold at 20 returns a TPR of 97.1% but a FPR of 25.67%, meaning that a fourth of the point considered accurate should actually not be included because they are inaccurate. However, FIG. 6D, presenting the same analysis as in FIG. 5 but only for the FPR points, shows that FPR points still have a RMSE error for SDNN below 20 ms that, depending on the analysis, can still be considered acceptable.

[0148] The present invention proposes to defined a quality metric not on the PPG waveform but on the HRV metrics, which is associated with the HRV accuracy. A higher HRV accuracy, lower HRV error, is associated with a better quality. Using the ECG signal to calculate HRV metrics that are considered reliable, avoids the problem of manually annotating the PPG signal 118, a tedious, subjective, process that could potentially results in erroneous labelling and misleading results. While the waves in the ECG signal are often clearly visible, these can be labelled with relative safety, the PPG waves are usually more complicated to assess. A quality measure based on a combination of several HRV metrics errors is more robust than a quality measure based on a individual HRV metric error. This quality is thus universal, in the sense that can be used for all HRV metrics and there is no need to estimate a quality for each individual metric. The more accurate the prediction of the HRV error, the lower the FPR error will be.

LIST OF REFERENCE NUMBERS

[0149] 110 portable photoplethysmogram device [0150] 112 illumination source [0151] 114 photodetector [0152] 116 ECG device [0153] 118 PPG signal [0154] 120 processing unit [0155] 121 feature extraction [0156] 122 box [0157] 124 evaluation step [0158] 126 ECG data [0159] 128 evaluation of the ECG data [0160] 130 determining at least one heart rate variability metric [0161] 132 determining at least one heart rate variability metric [0162] 134 Determining of heart rate variability error [0163] 136 determining the heart rate variability accuracy [0164] 138 determining at least one performance metrics [0165] 140 estimate heart rate variability error [0166] 142 binary decision [0167] 144 vertical line