Method for the real-time reconstruction of an ECG from a PPG signal

Abstract

The invention relates to a method for generating an ECG signal from a PPG signal and a trained machine learning model comprising: ■ a first acquisition (ACQ1) of a first signal (S1), referred to as PPG signal, by means of a sensor arranged on a device for acquiring the signal which is in contact with the skin of an individual in order to measure the variations in the pulsatile blood volume; ■ segmenting (SEG1) the first signal (S1); ■ using a trained machine learning model to produce a second segmented signal of an ECG from an input vector comprising the data from one cycle of a first PPG signal; ■ recombining the segments produced in order to reconstruct a continuous ECG signal based on: o detecting two consecutive ECG markers of the second signal present in each cycle of the second signal; o reconstructing the ECG signal by concatenating the segments from the time markers.

Description

[0001] METHOD FOR RECONSTRUCTING AN ECG IN REAL TIME FROM A PPG SIGNAL

[0002] Field of invention

[0003] The field of the invention relates to the field of methods and devices for transforming, analyzing and processing cardiac signals. In particular, the field of the invention relates to the field of methods for transforming a photoplethysmography signal, called a PPG signal, into an electrocardiographic signal, called an ECG signal, in order to produce a continuous ECG signal from simple acquisition equipment.

[0004] State of the art

[0005] Currently, there are solutions that allow obtaining an electrocardiograph from a photoplethysmogram.

[0006] Unlike PPG, the ECG has technical implementation disadvantages. To perform the ECG, it is necessary to use multiple electrodes, while the PPG uses a single sensor to perform a continuous recording. Furthermore, the material used to provide a good quality ECG signal with the electrode can cause skin irritation, discomfort, and even allergic reactions during long-term use of these devices.

[0007] Certain factors such as the positioning of the electrodes, their detachment or their retention on the skin can influence the results of the ECG and lead to false positive or negative results.

[0008] There are methods in the prior art for reconstructing an ECG signal from a PPG signal using a machine learning algorithm. In this type of method, a segmentation of the PPG signal of a fixed length with several epochs is used.

[0009] One problem is that an ECG signal contains many signal descriptors, namely data characterizing the QRS complex and heart rate data, as well as variations in this frequency over time. Learning does not allow for reliable restitution of both the QRS complex and heart rate variability.

[0010] Consequently, the accuracy of predictions from current models remains unsatisfactory for exploiting the ECG signal reconstructed from a PPG signal, particularly to provide assistance in the diagnosis of cardiac pathologies.

[0011] There is a need to improve the accuracy of reconstructing an ECG signal from a PPG signal to obtain a continuous signal and generate indicators characterizing singularities of the ECG signal thus reconstructed.

[0012] Summary of the invention

[0013] One objective of the invention is to overcome the aforementioned drawbacks.

[0014] According to a first aspect, the invention relates to a method for generating an ECG signal from a PPG signal and a trained machine learning model comprising:

[0015] ■ A first acquisition of a first signal, called the PPG signal, by means of a sensor arranged on a signal acquisition device in contact with the skin of an individual to measure variations in the pulsatile volume of the blood;

[0016] ■ A sampling of the first signal at a predefined frequency;

[0017] ■ A recording of the heart rate;

[0018] ■ Segmentation of the first signal comprising: o Definition of a first time window of a first predefined duration; o Extraction of a first set of sampled points of the first PPG signal, called PPG points, comprising a first subset of points between two consecutive PPG markers and a second subset of consecutive points of the second PPG marker, the number of points of the second subset being less than the number of bridges of the first subset; o Positioning of said first segmented signal in the first time window from the second PPG marker;

[0019] ■ Use of a trained learning model to produce an output vector defining a second segmented signal of an ECG from an input vector comprising the data of the first set; ■ Recombination of the segments produced to reconstruct a continuous ECG signal from: o Detection of at least two consecutive ECG markers of the second signal present in each cycle of said second signal; o Reconstruction of the ECG signal by concatenation of the segments from the time markers.

[0020] An advantage of using a time window with less than two cycles is that it provides a reliable and accurate QRS complex reconstruction method. Learning has very good performance and the reconstruction is of better quality than when using a plurality of cycles considered as input to the network.

[0021] According to one embodiment, the heart rate is used to identify the local maximum corresponding to the desired ECG marker.

[0022] According to one embodiment, the ECG markers used during segment recombination are R peaks. An advantage is to minimize the error during cycle segmentation.

[0023] According to one embodiment, two consecutive markers of the first signal are used to interpolate the lengths between two consecutive ECG markers of the second signal.

[0024] According to one embodiment, the machine learning model is trained according to the method for training a neural network to generate an ECG from a PPG signal according to the invention.

[0025] According to another aspect, the invention relates to a method for training a neural network to generate an ECG from a PPG signal, said method comprising:

[0026] ■ A first acquisition of a first signal, called the PPG signal, by means of a sensor arranged on a signal acquisition device in contact with the skin of an individual to measure variations in the pulsatile volume of the blood;

[0027] ■ A second acquisition of a second signal, called an ECG signal, by means of at least one pair of electrodes affixed to the skin of an individual;

[0028] ■ A sampling of the first signal and the second signal each at a predefined frequency; ■ A recording of the heart rate;

[0029] ■ A synchronization of the first signal with the second signal, said synchronization comprising: o Detection of at least one first PPG marker of the first signal present in each cycle of said first signal; o Detection of at least one first ECG marker of the second signal present in each cycle of said second signal; o Estimation of a transit time of the first PPG signal; o Alignment of the two signals from an offset of one of the two signals by the value of the transit time;

[0030] ■ Segmentation of the first signal and the second signal comprising: o Definition of a first time window of a first predefined duration; o Extraction of a first set of sampled points of the first PPG signal, called PPG points, comprising a first subset of points between two consecutive PPG markers and a second subset of consecutive points of the second PPG marker; o Positioning of said first segmented signal in the first time window, said window comprising the two consecutive PPG markers; o Definition of a second time window of a second predefined duration; o Extraction of a second set of sampled points of the second ECG signal, called ECG points, comprising a first subset of points between two consecutive ECG markers and a second subset of consecutive points of a second marker;o Positioning said second segmented signal in the first time window, said window comprising the two consecutive ECG markers;

[0031] ■ Training a machine learning model comprising: o Generation of input vectors comprising the first set of points; o Implementation of a loss function from the second set of points;

[0032] ■ Generation of a trained machine learning model.

[0033] According to one embodiment, the first marker is an R peak of a QRS complex and in that the second marker is a local minimum of the second signal.

[0034] According to one embodiment, the transit time is estimated by calculating the minimum difference measured between two markers of each of the two signals considered in the same cycle. One advantage is that it allows the PPG and ECG signals to be synchronized homogeneously for each cycle.

[0035] According to one embodiment, the second subset of points of the second consecutive segmented ECG signal of a second marker comprises a number of points less than the number of points of the first subset.

[0036] One advantage is that it allows for precise recalibration at the network output to heart rate. In particular, by selecting a marker that defines a local minimum or a local maximum, the second set of points makes it possible to better detect these minimums or maximums at the network output and therefore to group the segments together according to heart rate.

[0037] According to one embodiment, a third subset of points of the second set of points comprises a set of ECG points of the same value succeeding the second subset of points so as to complete the second time window. An advantage is to make it possible to obtain windows of fixed size for training by considering a signal that can be up to a predefined duration. An advantage of the filling is not to disturb the learning with data from the QRS complex of a cycle other than the cycle defining the input of the network.

[0038] According to one embodiment, the acquisition of the heart rate is carried out by means of the analysis of the second ECG signal comprising the estimation of the intervals between two consecutive peaks of the acquired ECG signal, called RR intervals. An advantage is to obtain an accurate heart rate. According to one embodiment, the method comprises an indicator of the deviation of the heart rate calculated from the PPG signal and that calculated from the ECG signal. An advantage is to take into account the deviation between the two measurements during learning.

[0039] The following characteristics apply to both the method for generating an ECG signal from a PPG signal and a trained machine learning model and the method for training the machine learning model.

[0040] According to one embodiment of one and / or the other of these methods, the second subset of points of the first segmented PPG signal consecutive to a second marker comprises a number of points less than the number of points of the first subset. An advantage is to allow precise recalibration at the output of the network to the heart rate. In particular, by selecting a marker which defines a local minimum or a local maximum, the second set of points makes it possible to better detect these minimums or maximums at the output of the network and therefore to group the segments together according to the heart rate.

[0041] According to an embodiment of one and / or the other of these methods, a third subset of points of the first set of points comprises a set of PPG points of the same value succeeding the second subset of points so as to complete the first time window. An advantage is to make it possible to obtain windows of fixed size for training by considering a signal which can be up to a predefined duration. An advantage of the filling is not to disturb the learning with data from the QRS complex of a cycle other than the cycle defining the input of the network.

[0042] According to one embodiment of one and / or the other of these methods, the machine learning model is a convolutional neural network known as CNN.

[0043] According to one embodiment of one and / or other of these methods, the CNN neural network is a U-NET type network composed of an encoder, a decoder and a tool for focusing the neural network on the most important or relevant parts of the data, a tool called Attention Gate. According to one embodiment of one and / or other of these methods, the training of the coefficients of the machine learning model is carried out using an optimizer and an implementation of a cost function.

[0044] According to one embodiment of one and / or other of these methods, the acquisition of the heart rate is carried out by means of the analysis of the first PPG signal comprising the estimation of the intervals between two consecutive markers of the acquired PPG signal. An advantage is to obtain the heart rate directly from the PPG sensor and a computer.

[0045] According to another aspect the invention relates to a device for generating a continuous ECG signal comprising a PPG signal sensor and a computer and a memory for implementing the method of the invention. An advantage is that it allows such a device to be placed on a bracelet with a single point of contact on the wrist.

[0046] Brief description of the figures

[0047] Other characteristics and advantages of the invention will emerge on reading the detailed description which follows, with reference to the appended figures, which illustrate:

[0048] ■ [Fig.1]: A flowchart of a system for preprocessing and training a machine learning model from two acquired signals including an ECG signal and a PPG signal;

[0049] ■ [Fig.2]: A flowchart of training a neural network;

[0050] ■ [Fig.3]: A detailed flowchart of the operation of the trained neural network in Figure 2;

[0051] ■ [Fig.4]: A flowchart of the operation of the trained neural network;

[0052] [Fig.5]: A flowchart of the preprocessing system, training the neural network;

[0053] ■ [Fig.6]: An example of a PPG signal acquired by a measuring device;

[0054] ■ [Fig.7]: The example of the PPG signal sampled at 100Hz after filtering;

[0055] ■ [Fig.8]: An example of raw ECG signal acquired by electrodes; ■ [Fig.9]: The example of ECG signal sampled at 100Hz after filtering;

[0056] ■ [Fig.10]: An example of filtered and aligned PPG and ECG signals with the R peaks of the ECG signal and the corresponding minima of the PPG signal;

[0057] ■ [Fig.11]: An example of segmentation of the PPG signal on a window comprising a portion of a following cycle and a filling;

[0058] ■ [Fig.12]: An example of segmentation of the ECG signal on a window comprising a portion of a following cycle and a filling;

[0059] ■ [Fig.13]: An example of segmentation with a fixed two-second window of the ECG signal;

[0060] ■ [Fig.14]: An example of segmentation with a fixed two-second window of the filtered PPG signal;

[0061] ■ [Fig.15]: An example of neural architecture of a machine learning algorithm model for automatic reconstruction of an ECG;

[0062] ■ [Fig.16]: an example of representation of the output signals of the network trained with a first type of segmentation;

[0063] ■ [Fig.17]: an example of representation of the output signals of the network trained with a second type of segmentation;

[0064] ■ [Fig.18]: an example of representation of the signals recombined by the method of the invention.

[0065] Detailed description

[0066] According to a first aspect, the invention relates to a method for generating an ECG signal from a PPG signal and an MLi machine learning model. This method is for example described in Figures 3 and 4.

[0067] According to a second aspect, the invention relates to a method for training an MLi machine learning model from the acquisition of an ECG signal and the acquisition of a PPG signal. This method implements a cost function to learn the model. This method is for example described in Figures 1 and 2.

[0068] Acquisition of the PPG signal The method for generating an ECG signal comprises a first step of acquisition ACQi of a PPG signal noted signal Si.

[0069] Acquisition can be performed using different devices. For example, the device can be a bracelet or a watch with at least one PPG sensor.

[0070] According to one embodiment, the PPG sensor comprises an emitter of a light source, such as a diode and a receiver such as a photodetector.

[0071] The optical measurement used to generate a PPG signal and derive heart rate from this technique may be known as "optical heart rate monitoring" or "OHR / OHRM". The acronym PPG may sometimes be used to refer to the technology, signal, or equipment. In the remainder of the description, a signal from this technology will be referred to as a PPG signal, and equipment such as a sensor will be referred to as a PPG sensor.

[0072] Photoplethysmography, or PPG, is a non-invasive optical method for analyzing blood volume changes in superficial tissues. This method relies on analyzing changes in light absorption in tissues. This method is used, for example, in oximeters to measure blood oxygen saturation, and in smartwatches and bracelets to calculate heart rate.

[0073] According to one embodiment, the PPG sensor is an optoelectronic sensor. It is composed of an emitter of a light source, such as a diode and a receiver such as a photodetector. The diode emits light which passes through different successive layers of the skin: the surface, the stratum corneum, the epidermis and the dermis. Part of the light is absorbed by the blood, part of the light is reflected. The photodetector receives the reflected light.

[0074] According to one embodiment, the transmitter and the receiver are placed next to each other, called reflection mode.

[0075] According to one embodiment, the photoreceptor detects variations in reflected light and converts them into an electrical signal.

[0076] A cardiac cycle can be divided into two phases based on blood flow. When the heart contracts, it expels blood into the vascular system; this is the so-called systole phase. This results in an increase in blood volume in the vessels. The so-called diastole phase is the phase of dilation of the heart. This phase results in a decrease in blood volume in the vessels. In the systole phase, the increase in blood volume results in an increase in the light absorbed by the blood and a decrease in the intensity of the transmitted light. In the diastole phase, the decrease in blood volume results in an increase in the intensity of the transmitted light. Thus, the signal acquired by the sensor includes a variation in a physical parameter measured during this cyclical phenomenon.

[0077] According to one embodiment, the PPG signal denoted Si is a signal whose amplitude is expressed in volts or in an arbitrary unit as a function of time. According to one embodiment, the heart rate can be deduced from the signal Si. According to one embodiment, a computer configured to detect characteristic markers of the PPG signal makes it possible to deduce the heart rate.

[0078] In an illustrative example, with reference to figure 6, we obtain the graphic representation of a raw PPG signal noted Si sampled at 100Hz after its acquisition by a device.

[0079] ECG acquisition

[0080] During the learning phase and the implementation of the training method, the method acquires an ECG signal and a PPG signal in order to train the MLi machine learning model. The ECG signal is denoted S2 and the PPG signal is denoted Si.

[0081] In the latter case, according to one embodiment, an ECG signal is acquired from an individual whose PPG signal is also acquired. In order to achieve efficient training, the ECG and PPG signals used to train the ML1 model are synchronized and acquired from the same individual. One advantage is to train the model with correlated PPG and ECG signals.

[0082] According to one embodiment, during ECG acquisition, the individual is at rest. According to another embodiment, the acquisition of the ECG signal is carried out during an effort. According to one embodiment, ECG signals are acquired in different configurations corresponding to different states of an individual. One interest is to obtain the most exhaustive training possible of all the states of cardiac activity of the individual. According to one example, the acquisition is carried out during a stress test for the generation of a stress electrocardiogram. In this context, the individual practices increasingly intense physical effort.

[0083] According to one embodiment, the acquisition of the ECG signal comprises the implementation of an arrangement of ten adhesive electrodes connected to an electrocardiograph by means of cables. For a more precise measurement, twelve or eighteen electrodes can be used.

[0084] The four electrodes, called peripheral electrodes, are distributed around each wrist and ankle. At least six electrodes, called precordial electrodes, are distributed in the region of the thorax located in front of the heart, called the precordial zone.

[0085] In a lighter mode, two electrodes are used to obtain at least a potential difference.

[0086] According to one embodiment, the electrodes are stuck to bare skin.

[0087] An electrocardiogram, or ECG, is a method for visualizing changes in the electrical current flowing through the heart over time. This method relies on recording and transcribing the electrical currents flowing through the heart during each cardiac contraction. The ECG provides information on the rhythm, rate, and conduction of the heart. It allows for the analysis of electrical signals generated and conducted by the heart muscle or the analysis of blood circulation.

[0088] An electrical current from the heart appears at a specific point at the top of the right atrium, a point called the sinus node. The generated current propagates throughout the heart. When the current passes through the heart's atria, it causes the heart muscle to contract. The conduction system maintains the heart rate within a specific range. Electrical propagation may be disrupted due to heart disease.

[0089] According to one embodiment, the electrical current generated at the sinus node is measured between two points on the surface of the body using the electrodes and is captured by the electrocardiograph.

[0090] According to one embodiment, the signal obtained is the ECG signal S2 is a signal whose amplitude is expressed in Volts as a function of time. In an illustrative example, with reference to Figure 8, the graphical representation of a raw ECG signal sampled at 100Hz is obtained after acquisition by an electrocardiograph.

[0091] The method for generating an ECG signal comprises a step of recording the heart rate. This heart rate can be deduced from the ECG signal S2.

[0092] According to another embodiment, a calculator configured to detect characteristic markers of the ECG signal makes it possible to deduce the heart rate.

[0093] According to one embodiment, the heart rates are deduced from the two signals S1 and S2. In the latter case, the measurement of the two heart rates from the two signals makes it possible to carry out measurement error corrections on one or the other of the signals.

[0094] Preferably, the heart rate is deduced from the PPG signal. An advantage is to train the ML1 model in the operating conditions of the method of the invention in which the ECG is not acquired since we seek to regenerate it from the PPG signal.

[0095] Filtering

[0096] The method for training the ML1 model includes a step of filtering and sampling the PPG signal Si and the ECG signal S2 each at a predefined frequency.

[0097] According to one embodiment, the filtering and sampling step performed during the execution of the training method are identical to those that will be performed during the execution of the method for generating an ECG signal. In this case, the filtering and sampling steps are only performed on the acquired PPG signal Si.

[0098] According to one embodiment, the PPG signal is filtered. The use of a bandpass filter between 40 and 400 Hz to filter the high and low frequencies eliminates artifacts related to the movement of the individual. The step of filtering the PPG signal is denoted FILT1 in Figure 1. The step of sampling the PPG signal is denoted ECH1 in Figure 1.

[0099] In an illustrative example, with reference to Figure 7, the raw PPG signal sampled at 100Hz of Figure 6 is filtered and resampled to 125Hz.

[0100] The filtered signal Si also noted SIF of figure 7 represents a signal in which the baseline is made stable over time. According to one embodiment, the ECG signal is filtered using a method called the "Tarvainen" method to reduce the shape effect. Indeed, during the recording of the ECG, certain phenomena such as the movement of the individual, the detachment of the electrodes, the induced effects of respiration, variations in skin impedances, etc. can cause an overall wave-like tendency affecting the electrical signal produced by the heart. This phenomenon is also known as baseline drift. The baseline or isoelectric line of the heart corresponds to the plot of the average amplitude of the signal smoothed over a given period of time.

[0101] This filtering is based on the use of a high-pass filter. In the case of the so-called "Tarvainen" method, finite impulse response filters are used. This means that the filter output reacts only for a finite time after the application of an input impulse. This is a method known as "distending." Detending a signal consists of eliminating or reducing an underlying trend in the collected data, the trend corresponding to variations in the average signal other than those generated by the cardiac electrical activity measured over time.

[0102] In an illustrative example, with reference to Figure 9, the ECG signal sampled at 100Hz in Figure 8 is filtered so as to eliminate the "trend" phenomenon. It is verified that this filtering makes it possible to normalize the average value of the signal over time, reducing or even eliminating its variations over time. The ECG signal filtering step is noted FILT2 in Figure 1. The ECG signal sampling step is noted ECH2 in Figure 1. Preferably, the sampling on signal Si will be chosen to be identical to the sampling of signal S2.

[0103] According to other embodiments, other types of filtering can be implemented such as filtering of the electromyogram, filtering of interference noise at a certain predefined frequency.

[0104] The SIF and S2F signals are then recorded in a memory to be processed to train the ML1 model. The processing can correspond to the shaping of the model's input vector, data enrichment, or data control. Indeed, some signals can be eliminated when the quality of the latter is not sufficient and it does not allow the ML1 model to be correctly trained. The signal recording step is noted ENR1 in Figure 1. Synchronization - Alignment

[0105] When training the MLi model, the acquisition of PPG and ECG signals is performed simultaneously, therefore the context for acquiring an ECG signal is the same as for acquiring a PPG signal. By context, we mean data of days, times, dates, individual, physical state of the individual, etc.

[0106] When the MLi model is trained and a first acquisition ACQi of a PPG signal is carried out jointly with a second acquisition ACQ2 of an ECG signal, according to one embodiment, the method for generating an ECG signal comprises a step of synchronizing the filtered ECG signal and the filtered PPG signal. This step is denoted SYNCi in Figure 1. This step is also called signal alignment.

[0107] This step has the advantage of improving the learning of the MLi machine learning model because the shifts and variations of these shifts over time do not alter the training of the neural network. This synchronization makes it possible to erase the longer transit times of the PPG signal with respect to the ECG signal given the nature of the signals and the body conducting them.

[0108] This synchronization phase is no longer carried out during the exploitation phase of the trained neural network given the fact that only the PPG signal will be acquired.

[0109] The graphic representation of an ECG signal groups together a set of waves broken down as follows:

[0110] ■ The P wave representing the contraction of the heart's atria

[0111] ■ The so-called "QRS" complex corresponds to mechanical systole, in other words to ventricular contraction. It contains three waves: the Q wave is the first negative deflection, the R wave is the first positive deflection and the S wave is the second negative deflection following the R wave.

[0112] ■ The T wave represents the so-called repolarization phase of the ventricles, in other words the return to the resting phase of the ventricles.

[0113] The Q, R and S peaks are the peaks of each corresponding wave. Figure 14 shows an example of an S2 signal with a QRS complex in which the P, R and T waves are represented. The graphical representation of a PPG signal groups together a set of waves broken down as follows:

[0114] ■ The systolic wave which reaches its maximum SY during cardiac systole at a point called “the peak” or “systolic summit”;

[0115] ■ The diastolic wave observed between the time at which the heart completes the systolic cycle and the diastolic peak DI.

[0116] A representation of these two waves is shown in Figure 11.

[0117] It is understood that ECG and PPG signals are cyclical signals due to the heart rate and due to the reproduction of a pattern at each period in time with the heart rate. These patterns make it possible to discriminate areas of interest allowing the definition of time markers in the cycle to synchronize the signals.

[0118] In order to achieve synchronization of these signals, different steps can be implemented.

[0119] According to one embodiment, the first step for synchronizing the ECG signal and the PPG signal comprises determining a first marker of the PPG signal and a second marker of the corresponding ECG signal. These markers are, for example, local minima or maxima such as characteristic peaks.

[0120] In one example, the markers used are the R peaks of the ECG signal and the trough corresponding to the R peak of the PPG signal taken in the same cycle. Figure 10 illustrates the markers MECGI and MECG2 which are consecutive R peaks on the signal Si. Figure 10 illustrates the markers MPPGI and MPPG2 which are consecutive troughs on the signal S2.

[0121] According to one embodiment, the second step comprises estimating the transit time of the PPG signal relative to the transit time of the ECG signal.

[0122] A pulse transit time is a measure of the transmission delay between the ECG signal and the PPG signal at a certain point in the body.

[0123] According to one embodiment, a measurement of the time difference between an ECG marker and a corresponding marker for the PPG is performed for each cycle. According to one embodiment, the transit time is considered as the minimum difference measured between two markers of each of the two signals considered from the same cycle. It is understood that the transit time is globally constant over time since this delay depends only on the physics of the signals propagating in a body. By considering the minimum difference, the method of the invention makes it possible to dissociate the transit time from all other phenomena altering the signal measurements.

[0124] In an illustrative example, with reference to Figure 10, the filtered ECG and PPG signals of Figures 7 and 9 are synchronized using as markers, the R peaks, for the ECG signal, and the corresponding troughs for the PPG signal.

[0125] Signal segmentation

[0126] ECG signal

[0127] The method for training the model of the invention comprises a step of segmentation SEGi of the ECG signal S2 and the PPG signal Si. According to one embodiment, the training method and the method for reconstructing an ECG implement a segmentation of each cardiac cycle of the PPG signal. The method for training the model ML1 implements a step of segmentation of the ECG signal S2 necessary for training the model. The segmented signals are denoted respectively Si' and S2' relative to the segmented PPG signals and the segmented ECG signals.

[0128] Figure 12 illustrates a segmentation of an ECG signal in an FTI window.

[0129] The duration of the segmentation window is advantageously strictly less than two cardiac cycles.

[0130] One advantage is to learn the model with data that essentially characterizes the waveform over one cycle and not with data that characterizes the rhythm or frequency of the signal.

[0131] This segmentation window has the advantage of improving the learning of the machine learning model because it mitigates the reconstruction error of the ECG signal obtained by machine learning for the “QRS” complex.

[0132] This advantage results from the fact that the training data only represents the signal data over a single cycle. In other words, in the case of segmentation of several cycles, the training will involve learning the reconstruction of the rhythm. However, in the present invention, the heart rate is obtained by an analysis of the ECG and / or PPG signal not necessarily using a machine learning model. This heart rate is then used to reconstruct or consolidate the reconstruction of the segmented signal into a continuous signal of all the cycles.

[0133] According to various embodiments, the segmentation window comprises at least one complete cycle and possibly a fraction of a following and / or preceding cycle. This window is identical when training the model and when using the learned model.

[0134] According to one embodiment, the method for training the model and the method for generating an ECG signal comprises a step of segmenting SEGi the signals.

[0135] Regarding the method for training the model, the latter comprises an extraction of a second set of sampled points E2 from the second ECG signal S2.

[0136] According to one embodiment, a time window FT2 of a predefined duration is fixed. For the ECG signal, this window FT2 is called the second time window. For the PPG signal, the equivalent segmentation window will be noted as the first time window, it is noted FTI.

[0137] We denote by E2 the set of points extracted from the ECG signal defining a segmented signal S2' resulting from the segmentation of the ECG signal in a time window FT2. The points of the set E2 make it possible to define an input vector of an ML1 machine learning model such as a neural network.

[0138] We denote by E1 the set of points extracted from the PPG signal defining a segmented signal Si' resulting from the segmentation of the PPG signal in an FTI time window in order to define an input vector of an ML1 machine learning model such as a neural network.

[0139] According to one embodiment, two consecutive ECG markers denoted MECGI, MECG2 are defined. These markers are defined so as to obtain a complete cycle. These markers make it possible to collect a first set of sampled points denoted ENS21 between said two markers MECGI and MECG2. These ENS21 points are extracted from the received or acquired signal.

[0140] A second subset of sampled points denoted ENS22 is extracted from the second ECG marker MECG2 for a predefined duration following the second predefined MECG2 marker. The sets of points ENS21 and ENS22 partly compose the set of points E2.

[0141] In one example, the consecutive markers used are two consecutive R peaks of the ECG signal S2'. They are denoted MECGI, MECG2 in Figure 10.

[0142] According to one embodiment, the size of the point set ENS22 is smaller than the size of the point set ENS21. This ensures that a complete cycle is obtained while strictly limiting the points to two cycles.

[0143] According to one embodiment, a third subset of points ENS23 of the second set of points E2 comprises a set of points of the same value succeeding the second subset of points ENS22 so as to complete the second time window FT2. The set of points ENS23 completes the set of points E2. Indeed, since the window FT2 has a fixed duration and the heart rate can change, the number of points taken from the sets ENS21 and ENS22 can vary and consequently, the size of the set ENS23 can vary. This third set of points makes it possible to perform a filling so as to maintain a fixed segmentation window.

[0144] In an illustrative example, with reference to Figure 12, this segmentation is used for the ECG signal of Figure 10. All points in the third subset of points ENS23 have a zero value.

[0145] PPG signal

[0146] The method for generating an ECG signal comprises a step of segmentation SEG1 of the PPG signal Si. This method comprises an extraction of a first set of sampled points E1 of the first PPG signal Si. The segmentation of the signal is preferably identical during the method for training the ML1 model.

[0147] Figure 11 illustrates a segmentation of a PPG signal in an FTI window.

[0148] According to one embodiment, a first FTI time window of a predefined duration is fixed. According to one embodiment, the FTI time window is of the same size as the second F T2 time window.

[0149] According to one embodiment, the method of the invention makes it possible to define two consecutive PPG markers denoted MPPGI, MPPG2. A first set of sampled points denoted ENS11 is extracted between the two markers MPPGI and MPPG2. A second subset of sampled points denoted ENSi2 is extracted from the second PPG marker MPPG2 up to a predefined point. The set of points Ei comprises the subsets of points ENS11 and ENS12.

[0150] According to an example, the two consecutive PPG markers MPPGI, MPPG2 are two consecutive minima corresponding substantially to the two markers MECGI, MECG2 of the ECG signal in the learning process of the MLi model. An interest is to retain points that are strongly correlated and correspond to the same physiological phenomena.

[0151] According to one embodiment, the length of the set of points ENS12 is less than the length of the set of points ENSn. This ensures that the window is strictly less than two epochs, in other words two cardiac cycles.

[0152] According to one embodiment, a third subset of points ENS13 of the second set of points Ei comprises a set of PPG points of the same value succeeding the second subset of points ENS12 so as to complete the first time window FTI. The set of points ENS23 completes the set of points ENS11 and ENSi2.

[0153] In an illustrative example, with reference to Figure 11, this segmentation is used for the PPG signal of Figure 10. The set of points in the third subset of points ENS has a zero value.

[0154] Other segmentation

[0155] According to a second embodiment, another segmentation is possible for the ECG and PPG signals. The segmentation takes all the points included in a predefined time interval. The segmentation window includes at least one cardiac cycle and can go beyond two cycles.

[0156] In an illustrative example, with reference to Figures 13 and 14, a segmentation with a two-second window is used for the PPG Si' and ECG S2' signals. The segmentation is 250 points long with 125Hz sampling.

[0157] In order to estimate the performance of the models, the relative root mean square error rRMSE and the Pearson correlation coefficient p value can be estimated to evaluate the methods. The first segmentation achieves very good results with p = 0.94 + / - 0.05 and RSME = 0.05 + / - 0.02. The second method also achieves good results but lower than the first segmentation method with p = 0.68 + / - 0.24 and RSME = 0.11 + / - 0.04

[0158] Pearson correlation is a statistical measure that assesses the linear relationship between two continuous variables. It describes the extent to which the variation in one variable is associated with the variation in another variable. The Pearson correlation takes its values ​​in the range between -1 and 1 . The closer the Pearson correlation coefficient is to -1 or 1 , the stronger the correlation between the two variables. Here, the ECG signal obtained by machine learning follows a similar trend to the expected ECG signal. That is, the two signals increase or decrease together in a linear manner.

[0159] Training a machine learning (MLi) model

[0160] Figure 2 represents the different signals processed during the training of the MLi model. The PPG signal acquired by a measuring device is denoted Si. The ECG signal acquired by another measuring device is denoted S2. The segmented signals processed by the neural network to be trained are denoted Si' and S2' respectively. The output signal of the neural network to be trained is denoted 82a'.

[0161] The trained ML1 machine learning model can predict the continuous ECG signal from a single PPG signal measurement. Training the model transforms the input vector from the PPG into an output vector that is as close as possible to the ECG signal that would have been measured by an ECG measuring device if it were used.

[0162] According to one embodiment, during the machine learning phase, the machine learning model receives as input vectors the set of sampled points E2 from the ECG and the set of sampled points E1 from the PPG.

[0163] According to one embodiment, the machine learning model CNN1 is a convolutional neural network called CNN.

[0164] According to an exemplary embodiment, the machine learning model comprises a U-NET type architecture. Such a U-NET architecture is itself based on a CNN architecture. In this case, the model comprises an ENC encoder, a DEC decoder, a software component for focusing the neural network on the most important or relevant parts of the data, a software component called “Attention Gate” and a memory layer represented by the “Biderectional Gated Recourent Unit” called BiGRU.

[0165] BiGRU is a type of recurrent neural network. It consists of two GRU blocks, one in the forward direction of data propagation and the other in the reverse direction.

[0166] In another embodiment, an LSTM layer may be used instead of the GRU block. However, this embodiment may require longer training.

[0167] According to one embodiment, three blocks are used to implement the ENC encoder. Each block includes at least one convolution layer. The convolution layer, denoted CONV in Figure 15, uses filters that scan the input data by performing convolution operations.

[0168] In one example, the core dimensions of each layer increase as the layers advance.

[0169] Each block also includes a normalization layer, such as a layer called "Layer normalization" in English terminology. This layer is denoted LN in Figure 15. It allows the outputs of each layer of the neural network to be normalized. This layer allows the stabilization of the neural network, to accelerate the learning of the model by allowing it to converge more quickly.

[0170] The latter allows faster and more stable learning thanks to the standardization of layer inputs.

[0171] Each block also includes an activation function, such as a rectified linear activation called ReLU, denoted LR. This function allows neurons to be activated or deactivated by adding biases. The so-called "ReLU" layer is also called a linear rectification unit. The activated features are passed on to the next layer, which promotes faster and more efficient learning.

[0172] Figure 15 represents an example of the architecture of a U-NET type network comprising 3 blocks in the ENC encoder and 3 blocks in the DEC decoder. The convolution layers are denoted CONV, the input normalization layers are denoted LN for "Layer normalization" in English terminology and the activation layers are denoted LR.

[0173] The second block and the third block are denoted RB and correspond to residual blocks. One advantage of a residual block is to improve learning in deep neural networks by allowing a more direct flow of the gradient during backpropagation.

[0174] The implementation of a residual block (RB) provides a skip connection, which bypasses one or more layers. This connection passes the input of the block directly to its output, summing it to the output of the intermediate layer(s). This is represented in Figure 15 in the RB block by an arrow starting from the input of the block and allowing the variables or coefficients to be added to the variables or coefficients at the output of the RB block before the last activation layer (LR). One benefit is to allow the error signal to propagate more efficiently through the network during backpropagation. This helps to mitigate the gradient vanishing problem, especially in very deep networks. A residual block (RB) includes at least one convolution layer (CONV), one normalization layer (LN), and one rectified activation layer (LR). Figure 15 represents an example of such an architecture.

[0175] According to one embodiment, two “Attention Gates” in English terminology and noted AG in figure 15 connect the encoder ENC and the decoder DEC.

[0176] According to one embodiment, the decoder also comprises three blocks. The first two blocks each respectively comprise a transposed convolution layer CONV_T, a normalization layer denoted LN and a rectified activation layer ReLU, denoted LR. The final block of the decoder comprises two transposed convolution layers CONV_T separated by a memory layer composed of a BiGRU denoted BG in Figure 15.

[0177] The DEC decoder allows the reconstruction of the segmented signal from the characteristics learned by the encoder.

[0178] Cost function According to one embodiment, the training method comprises the implementation of a cost function making it possible to calculate the error between the produced output 82a' representing the reconstructed ECG signal and the initial measured ECG signal S2'. The machine learning model generates an ECG signal 82a' which is therefore compared to the signal S2' corresponding to the input E2 to calculate this error and minimize it by modifying the coefficients of the convolution layers. The difference A = |S2 -S2b'| is measured from a distance function such as a quadratic error for example.

[0179] According to one embodiment, the method comprises a feedback loop, for example carried out by means of gradient descent, which allows the network to be trained by modifying the coefficients of the matrices of the convolution layers.

[0180] The ECG signals measured by the S2 electrodes can be split into two groups. A first set of signals is used within the cost function for network training and the second set of signals is used to validate that sufficient training allows to obtain given performances.

[0181] The output signals of the neural network for training are denoted 82a' and for using the trained network the output signals of the network are denoted 82b'.

[0182] Recombination of signals to form the continuous ECG

[0183] Figure 3 represents the recombination steps of the reconstructed segmented signals. The recombination step is denoted RECOMB1. In this figure the segmentation step is denoted SEG1 and the processing step by the trained neural network is denoted ML1.

[0184] The PPG signal acquired by a measuring device is denoted Si. The signal processed by the trained neural network is denoted Si'. The signal output by the trained neural network is denoted 82b' and the recombined signal is denoted S2b.

[0185] Figure 16 shows 82b' ECG signals produced for different epochs following one another with the first segmentation. The signals S2b(tj)', S2b(ti+i)' and S2b(ti+2)' are shown.

[0186] Figure 17 shows ECG signals 82b' produced for different epochs following one another with the second segmentation. The signals S2b(tj)', S2b(ti+i )' and S2b(ti+2)' are shown with this second segmentation. With the first or second segmentation, the invention makes it possible to recombine these signals with each other in order to produce a continuous signal.

[0187] We denote S2b' the non-recombined signals at the output of the neural network and we denote S2b the recombined signals forming a continuous signal comprising a plurality of epochs.

[0188] For this purpose, the method for generating an ECG comprises a step of combining the signals formed by the output vectors of the neural network. The reconstruction is carried out continuously so as to form a continuous ECG signal. The sequence of reconstructed ECG signals is recorded in a memory so as to allow exploitation by a computer embedded in the PPG measuring device or by a remote computer, for example that of a computer, a digital tablet or a server.

[0189] The training method made it possible to generate segmented ECG signals faithful to those which could have been measured by electrodes of an ECG. An advantage of the method of the invention is that a simple PPG measuring device makes it possible to generate a continuous ECG signal which can be used to detect singularities of the ECG signal such as arrhythmias, QRS complex fragmentation rates, variations in heart rate.

[0190] In order to reconstruct the signals produced by the machine learning model, the method for generating the ECG comprises a detection of consecutive ECG markers. Figure 16 represents only the markers MECGI and MECG2 on the first signal S2b(ti). The markers used can be local minima or local maxima such as the marker MECGI and the marker MECG2. The second signals S2b generated by the machine learning model MLi will be "glued together" to form a continuous signal using the markers present in each cycle of said second signal S2'. In other words, the recombination can be carried out by concatenating the segments produced by the neural network from the temporal markers.

[0191] Marker detection can be performed using an algorithm that takes two elements into account: the search for local minimums or local maximums and the heart rate. Indeed, by correlating these two pieces of information, the method of the invention ensures that the correct local minimum is considered in combination with the heart rate analyzed upstream of the neural network, during signal acquisition.

[0192] With respect to the training method, the heart rate can be deduced from the PPG signal or the acquired ECG signal. With respect to the method of generating the ECG signal, the heart rate is advantageously deduced from the PPG signal acquired by the device.

[0193] One interest is to search for a local minimum or maximum around the heart rate to avoid considering a marker corresponding to the marker of the previous cycle. For this purpose a predefined margin of error can be used to search for a local maximum or minimum around a point in the cycle.

[0194] The segments are then combined so as to join one end of one segment with another end of the next segment instead of the maximum value of the local maximum or local minimum, that is, the marker generated at each cycle. For example, the segments can be joined together at the top of the R peaks.

[0195] Sometimes, because the segmented data of the ECG and the PPG do not perfectly conform to the relationship that exists between the PPG and the ECG signal, in particular because the arrival time of the pulse has been considered constant, the method of the invention comprises a step of interpolating the RR segments to the length of the local maxima or local minima of the corresponding PPG signal. One advantage is to reinject the known knowledge of the acquired PPG signal in order to obtain the most reliable and faithful reconstructed ECG signal possible.

[0196] Figure 18 represents a first diagram comprising the acquired PPG signal Si, a second diagram comprising the reconstructed ECG signal S2b, and a third diagram representing the two signals Si and S2b allowing the local maxima and local minima of each cycle to be visualized.

[0197] An advantage of the invention is that it allows an ECG to be reconstructed in real time and continuously from the measurement of a PPG signal. One advantage is that a PPG device only requires one measurement point and can be embedded in a device such as a bracelet or a watch.

[0198] An advantage of the invention is to take advantage of a trained neural network to reconstruct an ECG signal from a cycle of a PPG signal and to reconstruct the signal continuously through the analysis of the heart rate. Thus, the network's training is not affected by the frequency of the cycles, the variation of the rate and the possible errors related to the reconstruction of the heart rate by the neural network.

Claims

CLAIMS 1. Method for generating an ECG signal from a PPG signal and a trained machine learning (MLi) model comprising: ■ A first acquisition (ACQi) of a first signal (Si), called PPG signal, by means of a sensor arranged on a signal acquisition device in contact with the skin of an individual to measure the variations in the pulsatile volume of the blood; ■ A sampling (ECHi) of the first signal (Si) at a predefined frequency; ■ A recording (ENRi) of the heart rate (Fc); ■ Segmentation (SEGi) of the first signal (Si) comprising: o Definition of a first time window (FTI) of a first predefined duration (Ti); o Extraction of a first set of sampled points (Ei) of the first PPG signal (Si), called PPG points, comprising a first subset of points (ENSu) between two consecutive PPG markers (MPPGI, MPPG2) and a second subset (ENS12) of consecutive points of the second PPG marker (MPPG2), the number of points of the second subset (ENS12) being less than the number of bridges of the first subset (ENSu); o Positioning of said first segmented signal (Si') in the first time window (FTI) from the second PPG marker (MPPG2); ■ Use of a trained learning model to produce an output vector (VECG) defining a second segmented signal (S2) of an ECG from an input vector (VPPG) comprising the data of the first set (E1); ■ Recombination of segments (VECG) produced to reconstruct a continuous ECG signal from: o Detection of at least two consecutive ECG markers (MECGI) of the second signal (S2) present in each cycle of said second signal (S2); o Reconstruction of the ECG signal (S2b) by concatenation of segments from time markers.

2. Method for generating an ECG signal from a PPG signal and a machine learning model (MLi) trained according to the preceding claim, characterized in that the heart rate (F c ) is used to identify the local maximum corresponding to the desired ECG marker.

3. Method for generating an ECG signal from a PPG signal and a machine learning model (MLi) trained according to the preceding claim, characterized in that the ECG markers used during the recombination of the segments are R peaks.

4. Method according to the preceding claim, characterized in that two consecutive markers (MPPGI, MPPG2) of the first signal (Si) are used to interpolate the lengths between two consecutive ECG markers of the second signal (S2b).

5. Method according to any one of the preceding claims, characterized in that the machine learning model (MLi) is trained according to claim 6.

6. Method for training a neural network to generate an ECG from a PPG signal, said method comprising: ■ A first acquisition (ACQi) of a first signal (Si), called PPG signal, by means of a sensor arranged on a signal acquisition device in contact with the skin of an individual to measure the variations in the pulsatile volume of the blood; ■ A second acquisition (ACQ2) of a second signal (S2), called ECG signal, by means of at least one pair of electrodes placed on the skin of an individual; ■ A sampling (ECH1, ECH2) of the first signal (Si) and the second signal (S2) each at a predefined frequency; ■ A recording (ENR1) of the heart rate (Fc); ■ A synchronization (SYNCi) of the first signal (Si) with the second signal (S2), said synchronization (SYNCi) comprising: o Detection of at least one first PPG marker (MPPGI ) of the first signal (Si) present in each cycle of said first signal (Si); o Detection of at least one first ECG marker (MECGI ) of the second signal (S2) present in each cycle of said second signal (S2); o Estimation of a transit time (Ttr) of the first PPG signal (S2); o Alignment of the two signals (Si , S2) from an offset of one of the two signals (Si, S2) of the value of the transit time (Ttr); ■ Segmentation (SEG1) of the first signal (Si) and of the second signal (S2) comprising: o Definition of a first time window (FTI) of a first predefined duration (T1); o Extraction of a first set of sampled points (E1) of the first PPG signal (Si), called PPG points, comprising a first subset of points (ENSu) between two consecutive PPG markers (MPPGI, MPPG2) and a second subset (ENS12) of consecutive points of the second PPG marker (MPPG2); o Positioning of said first segmented signal (Si') in the first time window (FTI), said window comprising the two consecutive PPG markers (MPPGI, MPPG2); o Definition of a second time window (FT2) of a second predefined duration (T2);o Extraction of a second set of sampled points (E2) from the second ECG signal (S2), called ECG points, comprising a first subset of points (ENS21) between two consecutive ECG markers (MECGI, MECG2) and a second subset (ENS22) of consecutive points from a second marker (MECG2); o Positioning of said second segmented signal (S2) in the first time window (FTI), said window comprising the two consecutive ECG markers (MECGI, MECG2); ■ Training a machine learning model (ML1) comprising: o Generation of input vectors comprising the first set of points (E1); o Implementation of a loss function from the second set of points (E2); ■ Generation of a trained machine learning (ML1) model.

7. Method according to claim 6 characterized in that the first marker (M1) is an R peak of a QRS complex and in that the second marker (M2) is a local minimum of the second signal.

8. Method according to claim 6 characterized in that the transit time is estimated by means of calculating the minimum difference measured between two markers (Mi, M2) of each of the two signals (Si, S2) considered in the same cycle.

9. Method according to any one of claims 6 to 8 characterized in that the second subset (ENS22) of points of the second segmented ECG signal (S2) consecutive to a second marker (MECG2) comprises a number of points less than the number of points of the first subset (ENS21).

10. Method according to any one of claims 6 to 9 characterized in that a third subset of points (ENS23) of the second set of points (E2) comprises a set of ECG points of the same value following the second subset of points (ENS22) so as to complete the second time window (FT2).

11. Method according to any one of claims 6 to 10, characterized in that the acquisition of the heart rate (Fc) is carried out by means of the analysis of the second ECG signal (S2) comprising the estimation intervals between two consecutive peaks (Ri) of the acquired ECG signal, called RR intervals.

12. Method according to claim 11 characterized in that it comprises an indicator of the difference between the heart rate (Fc) calculated from the PPG signal (Si) and that calculated from the ECG signal (S2).

13. Method according to any one of claims 1 to 12 characterized in that the second subset (ENS12) of points of the first segmented PPG signal (Si') consecutive to a second marker (MPPG2) comprises a number of points less than the number of points of the first subset (ENS21).

14. Method according to any one of claims 1 to 13 characterized in that a third subset of points (ENS13) of the first set of points (E1) comprises a set of PPG points of the same value succeeding the second subset of points (ENS12) so as to complete the first time window (FTI).

15. Method according to any one of claims 1 to 14 characterized in that the machine learning model (CNN1) is a convolutional neural network called CNN.

16. Method according to any one of claims 1 to 15, characterized in that the CNN neural network is an II-NET type network composed of an encoder, a decoder and a tool making it possible to concentrate the neural network on the most important or relevant parts of the data, a tool called Attention Gate.

17. Method according to any one of claims 1 to 16, characterized in that the training of the coefficients of the machine learning model (ML1) is carried out using an optimizer and an implementation of a cost function.

18. Method according to any one of claims 1 to 17, characterized in that the acquisition of the heart rate (Fc) is carried out by means of the analysis of the first PPG signal (Si) comprising the estimation of the intervals between two consecutive markers of the acquired PPG signal.

19. Device for generating a continuous ECG signal comprising a PPG signal sensor and a computer and a memory for implementing the method of any one of claims 1 to 5.

Images