Data processing and blood pressure estimation techniques

Abstract

A data processing method comprising at least one signal that forms a plurality of patterns, for example, a photoplethysmographic signal whose patterns are representative of an individual's heartbeats, comprising comparing several of these patterns with each other and selecting the patterns that satisfy at least one predetermined similarity criterion. A method for training an artificial intelligence model using training data formed from data processed by such a data processing method. A method for estimating a physiological parameter, such as a blood pressure value, for one or more individuals, using an intelligent model trained with such a training method. A device and related computer program.

Description

Title of the invention: Techniques for data processing and blood pressure estimation technical field

[0001] The invention relates to the field of signal and data processing, particularly physiological.

[0002] The invention is of particular, but not limiting, interest for the estimation of blood pressure, in particular from a signal obtained by photoplethysmography or from another type of cardiac signal. Prior art

[0003] Several techniques have been proposed in recent years to estimate the blood pressure of individuals from photoplethysmographic (PPG) signals, notably using artificial intelligence models.

[0004] The inventors have observed that known techniques produce frequent estimation errors and that the accuracy of the estimates depends strongly on the individuals. Description of the invention

[0005] The invention aims to overcome all or part of the drawbacks of prior art techniques.

[0006] According to a first aspect of the invention, it relates to a method for processing data, in particular physiological data, also called "data to be processed". "

[0007] In general, the processing method of the invention aims to create usable data from such data to be processed.

[0008] The data to be processed include at least one signal which forms a plurality of patterns.

[0009] By way of non-limiting example, at least one signal is a signal photoplethysmographic patterns whose patterns are representative of an individual's heartbeats.

[0010] According to the invention, the treatment process comprises: - a comparison of several of these motifs with each other, and - a selection of patterns that meet at least one predetermined similarity criterion.

[0011] In other words, the pattern or patterns not meeting at least one similarity criterion are not selected and are therefore not used to form the said usable data.

[0012] Such a selection makes it possible to exploit only patterns likely to contain useful information. Indeed, in a signal containing patterns, at least some of which are representative of the same phenomenon—in this example, a heartbeat—a pattern that is not similar to patterns actually carrying useful information—the latter being, in this example, related to a heartbeat—is probably a pattern that does not contain useful or exploitable information.

[0013] In one embodiment, at least one similarity criterion is chosen from a list including: - a correlation coefficient greater than a predetermined value, - a difference between a value of at least one pattern parameter and a value predetermined less than a predetermined gap, at least one parameter being able to be chosen from a list including a peak time, a systolic peak time, a diastolic peak time, an interval between the systolic peak time and the diastolic peak time and a dicrotic notch time.

[0014] This list is of course not exhaustive and may include other similarity criteria, for example using a machine learning technique such as data partitioning.

[0015] In one embodiment, the treatment process comprises: - a construction of average patterns, each based on at least two of the selected patterns, - a comparison of the average patterns with each other, - a selection of average patterns if they meet at least one predetermined similarity criterion.

[0016] Without limitation, at least one similarity criterion used to select average patterns may be chosen from the list indicated above, i.e., from a list including: - a correlation coefficient greater than a predetermined value, - a difference between a value of at least one pattern parameter and a value predetermined less than a predetermined gap, at least one parameter being able to be chosen from a list including a peak time, a systolic peak time, a diastolic peak time, an interval between the systolic peak time and the diastolic peak time and a dicrotic notch time.

[0017] In other words, the average pattern(s) not meeting at least one similarity criterion are not selected and are therefore not used to form the said usable data.

[0018] In one embodiment, said plurality of patterns of at least one signal are patterns selected from initial patterns respecting at least one predetermined quality criterion.

[0019] The initial pattern(s) not meeting at least one quality criterion are not selected and therefore do not constitute said plurality of patterns which are the subject of the aforementioned steps, in particular the selection based on at least one similarity criterion.

[0020] In one embodiment, at least one quality criterion is chosen from a list including: - a signal-to-noise ratio lower than a predetermined value, - a difference between a value of at least one pattern parameter and a predetermined value less than a predetermined difference, the at least one parameter being able to be chosen from a list including a peak time, a systolic peak time, a diastolic peak time, an interval between the systolic peak time and the diastolic peak time, a dicrotic notch time, an instantaneous heart rate, an instantaneous heart rate variation, a pattern amplitude and an amplitude difference between a pattern start time and a pattern end time.

[0021] This list is of course not exhaustive and may include other quality criteria, for example a criterion relating to similarity with respect to a predetermined reference pattern.

[0022] In one embodiment, said at least one signal comprises several signals and said physiological data comprises one or more groups each comprising several data sets, each of these data sets comprising one respective of said signals as well as at least one blood pressure value.

[0023] Without limitation, the at least one blood pressure value may include both a systolic blood pressure value and a diastolic blood pressure value, or only one of these values.

[0024] In one embodiment, the processing method includes a selection of the group or groups in each of which the blood pressure values ​​of the datasets in that group are within a corresponding predetermined range of values ​​and / or exhibit variability less than a predetermined value.

[0025] In other words, the other group(s) are not selected and the signals they contain therefore do not form the said plurality of patterns which are the subject of the aforementioned steps, in particular the selection based on at least one similarity criterion.

[0026] In one embodiment, the processing method includes an operation of collecting said data.

[0027] The collection operation preferably includes, for each individual or group of individuals, the acquisition of several groups of datasets, each dataset comprising: - a signal that forms a plurality of patterns, preferably a photoplethysmographic signal whose patterns are representative of the heartbeats of that individual, - at least one blood pressure value for that individual, for example a systolic blood pressure value and / or a diastolic blood pressure value.

[0028] It is preferred that the signals be acquired consecutively within each of the groups.

[0029] The term "consecutive" means that the time interval between two successive datasets from the same group is relatively short compared to the time interval between two successive datasets belonging to different groups. For example, the time interval between two signals acquired successively within the same group may be on the order of a few seconds or minutes, while the time interval between two signals acquired successively within two different groups may be on the order of several hours.

[0030] According to a second aspect, the invention relates to a method for training an artificial intelligence model, also referred to as a "model" in this document.

[0031] The model training is preferably carried out using training data which are formed from data processed by a processing method as defined above.

[0032] Alternatively, the training data can be trained using any other method.

[0033] In one embodiment, the training data is formed from data collected using a collection operation similar to that described above.

[0034] The data collected may thus include, for each individual among one or more individuals, several groups of datasets, each of the datasets comprising: - a signal that forms a plurality of patterns, preferably a photoplethysmographic signal whose patterns are representative of the heartbeats of that individual, - at least one blood pressure value for that individual, for example a systolic blood pressure value and / or a diastolic blood pressure value.

[0035] It is preferred that the signals were acquired consecutively within each of the groups.

[0036] In one embodiment, the data are used such that: - a first input to the model receives data obtained based on the signal and at least one blood pressure value from one of the datasets of one of the individuals, this dataset being called the personalization set, - a second input of the model receives data obtained on the basis of the signal from another of said datasets of this individual, this other dataset being called the test set.

[0037] In one embodiment, during training, an output of the model is formed by a difference between at least one blood pressure value from the test set and at least one blood pressure value from the customization set.

[0038] In one embodiment, a third input of the model receives data obtained from data in the test set and / or the customization set.

[0039] In one embodiment, the data provided to one or more of said model inputs include two-dimensional data which can, for example, be obtained by time-frequency transformation of one or more of said signals and / or patterns.

[0040] Such a training method allows the model to achieve learning which inherently includes personalization by individual.

[0041] A single model thus trained can therefore be used to correctly estimate a blood pressure value with input data from different individuals.

[0042] According to a third aspect, the invention relates to a method for estimating a physiological parameter such as a blood pressure value of one or more individuals, using an artificial intelligence model trained with a training method as defined above.

[0043] Of course, the various processes described above are preferably implemented by computer.

[0044] Furthermore, the various processes described above are independent of each other and can each be implemented autonomously, optionally in combination with other techniques not described in this document.

[0045] The various processes of the invention can of course be combined with each other, the invention also relating to a method implementing a processing process as described above and / or a training process as described above and / or an estimation process as described above.

[0046] According to a fourth aspect, the invention relates to a device comprising a processing unit configured to implement a process according to any one of aspects defined above, preferably to implement all the processes described above.

[0047] In one embodiment, the device includes at least one sensor for measuring at least one physiological signal and / or at least one physiological parameter.

[0048] By way of example, at least one sensor may include a sensor for measuring a heartbeat signal, for example a piezoelectric sensor or a pressure sensor or a sensor comprising a camera capable of collecting a stream of images of an individual's face in order to generate a photoplethysmographic signal.

[0049] The at least one sensor may also include a sphygmomanometer-type sensor, for measuring an individual's blood pressure.

[0050] According to a fifth aspect, the invention relates to a computer program comprising executable computer instructions which, when executed by computer, implement a process according to any one of the aspects defined above, preferably several and more preferably all of the processes described above and the steps they comprise.

[0051] The computer program can be in any computer language, for example in machine language, C, C++, JAVA, Python, etc.

[0052] Other advantages and features of the invention will become apparent from the following detailed, non-limiting description. Brief description of the drawings

[0053] The following detailed description refers to the attached drawings on which: - [Fig.1] is a schematic view of a method according to the invention, comprising a phase of constructing training data, a phase of training an artificial intelligence model using the training data, a phase of constructing input data and a prediction phase in which the input data is provided to the trained model; - [Fig.2] is a schematic view of a data processing method according to the invention, this method being able to be implemented in the construction phase of the method of [Fig.1]; - [Fig.3] is a schematic view of a device according to the invention, this device being able to be used to implement all or part of the method of [Fig.1]. Detailed description of implementation methods

[0054] In the following non-limiting description, several innovative techniques are combined with the non-limiting objective of estimating a voltage value arterial of one or more individuals. These techniques can be implemented independently of each other, in applications and / or processes and / or devices different from those described here.

[0055] Fig. 1 schematically illustrates a method according to the invention.

[0056] Without limitation, this method implements a conventional artificial intelligence model of the regressive neural network type, in this example a deep learning convolutional neural network.

[0057] In general, the method in [Fig. 1] comprises a phase 1 of constructing training data, a phase 2 of training the model using the training data, a phase 3 of constructing input data, and a phase 4 of prediction in which the input data is provided to the trained model. These different phases and / or the steps and / or operations they contain can each constitute or be the subject of a respective process that can be implemented separately.

[0058] In the example of [Fig.1], the construction phase 1 includes an operation 6 of data collection, an operation 7 of pre-processing the collected data, an operation 8 of processing the pre-processed data, and an operation 9 of evaluating the data thus processed.

[0059] The collection operation 6 includes an acquisition of N sets of datasets, called "collected datasets", each of the sets comprising data from a respective individual.

[0060] The number N of series, and therefore of individuals, is generally greater than one, preferably greater than two, and more preferably greater than ten or one hundred. In this example, N is equal to 1000.

[0061] For each individual, the number M of data sets forming the series associated with that individual is generally greater than one, preferably greater than two and more preferably on the order of ten or hundreds.

[0062] In this example, each dataset in each series comprises: - a cardiac signal comprising patterns, also called impulses or beats, each corresponding to a respective heartbeat of the corresponding individual, - a systolic blood pressure reading for this individual, - a diastolic blood pressure value for this individual.

[0063] For each of the data sets of each of the series, the cardiac signal can correspond to a signal having a predetermined duration T, for example a duration of 30 seconds, acquired for example using a photoplethysmography (PPG) technique, thus forming a photoplethysmographic signal.

[0064] In this example, for each of the data sets of each of the series, systolic blood pressure and diastolic blood pressure are measured using a conventional sphygmomanometer during the acquisition of the corresponding cardiac signal.

[0065] In the example of [Fig. 1], data collection 6 is carried out so as to acquire, for each individual, a series of datasets forming several groups, each group comprising a number L of datasets acquired consecutively, that is, by defining a relatively short time interval between two successive datasets of the same group. For example, this intra-group time interval can be on the order of ten seconds between two signals. By comparison, the data of two successive groups of the same series are acquired by defining a relatively long inter-group time interval, for example, on the order of four hours between two signals.

[0066] In this example, for each of the groupings of each of the series of data sets, the number L of data sets of this grouping is generally greater than one, preferably greater than two, for example equal to three.

[0067] Thus, in this example, each of the series of data sets comprises a number of groupings equal to M / L.

[0068] As an indication, all the data for each of the series can be acquired over an overall period which can be several weeks or months, for example five weeks.

[0069] The preprocessing operation 7 generally includes one or more preprocessing steps of the cardiac signal from each of the collected data sets.

[0070] In this example, the preprocessing step(s) are chosen from a list including: - an interpolation and resampling step, - a filtering step, typically including windowing, for example using a Hamming window, - a differentiation step, by applying for example a second derivative.

[0071] Advantageously, the preprocessing operation 7 implements, for the cardiac signal of each of the collected datasets, the three steps listed above, preferably carried out in the order in which they are listed.

[0072] Such preprocessing makes it possible to remove noise from cardiac signals and more generally to clean them up in order to improve the detection of useful information they contain.

[0073] Datasets containing the signals thus pre-processed during operation 7 are called "pre-processed datasets".

[0074] The processing operation 8 in this example comprises a succession of steps for selecting and / or cleaning the pre-processed datasets, including but not limited to steps 11, 12, 13, 14 and 15 described below and schematically represented in [Fig.2].

[0075] Step 11 is carried out on the basis of the blood pressure values ​​from the datasets of each of the series.

[0076] Step 11 includes, for each of the data set groups of each of the series, a selection of the groups in each of which the blood pressure values ​​of the corresponding data sets meet one or more blood pressure quality criteria.

[0077] In this example, a first blood pressure quality criterion relates to the absolute value of the blood pressure values ​​and a second blood pressure quality criterion relates to the variability of these values.

[0078] More specifically, within the framework of this non-limiting example, step 11 includes, for each of the dataset groupings of each series, a selection of all the datasets in that grouping if: - the systolic blood pressure value of at least one of the datasets in this group is within a predetermined range of values, for example in the range 80-180 mmHg, and / or - the diastolic blood pressure value of at least one of the datasets in this group is within a predetermined range of values, for example in the range 40-120 mmHg, and / or - the variability of systolic blood pressure values ​​in the datasets of this group is less than a predetermined value, for example less than 5 mmHg, and / or - the variability of diastolic blood pressure values ​​in the datasets of this grouping is less than a predetermined value, for example less than 5 mmHg.

[0079] The datasets of the groups not selected during step 11 are not used for the rest of the processing.

[0080] Step 12 is carried out on the basis of the pre-processed signals, in particular the signals from the datasets selected during step 11.

[0081] By way of non-limitation, step 12 includes, for each of these signals, a selection of the dataset containing this signal if the patterns of this signal are representative of a heart rhythm having a value in a predetermined range of values, for example in the range 40-180 beats per minute.

[0082] To carry out this selection, one or more other criteria may be used in an alternative or complementary manner, for example a criterion based on an instantaneous variation in heart rate.

[0083] Data sets not selected during step 12 are not used in the rest of the processing.

[0084] Step 13 is carried out on the basis of the pre-processed signals belonging to the datasets that were selected during step 12.

[0085] Step 13 includes, for each of the patterns of each of these signals, a selection of this pattern if it meets one or more signal quality criteria.

[0086] In this example, a first signal quality criterion relates to a fluctuation and / or a signal-to-noise ratio of the part of the signal containing this pattern and a second signal quality criterion relates to one or more pattern parameters.

[0087] The pattern parameters can be chosen from a list including, but not limited to, a peak time, a systolic peak time, a diastolic peak time, an interval between the systolic peak time and the diastolic peak time, a dicrotic notch time, an instantaneous heart rate, an instantaneous heart rate variation, a pattern amplitude and an amplitude difference between a pattern start time and a pattern end time.

[0088] More specifically, within the framework of this non-limiting example, step 13 includes, for each of the patterns of each of these signals, a selection of that pattern if: - the signal-to-noise ratio of the portion of the signal containing this pattern is less than a predetermined value, for example less than 0.9 for a ratio that can fall within a range having a minimum value of 0 and a maximum value of 1, and / or - a value of one or more of said parameters of this pattern has a deviation less than a predetermined deviation from a corresponding predetermined value, for example if the absolute difference in amplitude between pattern start time and pattern end time is greater than 0.4 for a maximum amplitude normalized to 1, or if the systolic peak time is greater than 0.4 times the difference between pattern start time and pattern end time.

[0089] Step 14 is carried out on the basis of the pre-processed signals in which several patterns were selected in step 13. Datasets whose signal does not contain any pattern selected in step 13, or a number of patterns selected that is too small, for example less than two, are not used in the rest of the processing.

[0090] Step 14 includes, for each of these signals, a selection of patterns respecting one or more similarity criteria with respect to one or more other patterns of this signal.

[0091] Similarity criteria can be based on a correlation and / or a difference in absolute values.

[0092] In this particular example, for each of the patterns of each of these signals, step 14 includes a selection of that pattern if: - the correlation coefficient of this pattern with one or more other patterns of this signal is greater than a predetermined coefficient, for example greater than 0.95, and / or - the absolute difference between the value of at least one parameter of this pattern and the value of the same parameter of one or more other patterns of this signal is less than a corresponding predetermined value, for example a difference in systolic peak time of less than 7 ms.

[0093] Step 15 is carried out on the basis of the pre-processed signals in which several patterns were selected in step 14. Datasets whose signal does not contain any pattern selected in step 14, or a number of patterns selected that is too small, for example less than four, are not used in the rest of the processing.

[0094] Generally, step 15 includes, for each of these signals, processing this signal to form average patterns and comparing these average patterns in order to select this signal if the average patterns are sufficiently similar.

[0095] In this particular example, step 15 includes, for each of these signals, a calculation of the number of patterns in several parts of that signal, each of these signal parts being able to correspond to a fraction 1 / P of that signal, P preferably being an integer greater than or equal to two, for example, equal to two. In this example, this signal is selected if the number of patterns in at least one of these parts is greater than a predetermined value, for example, greater than one.

[0096] For each of the selected signals, step 15 implements signal processing consisting of normalizing the patterns of this signal and correcting their shape.

[0097] By way of non-limitation, shape correction may consist of modifying the signal so that each pattern has an identical amplitude between the pattern start time and the pattern end time.

[0098] For each of the parts of the signal thus processed, an average pattern is then calculated.

[0099] An average pattern may, for example, correspond to a pattern in which each of the values corresponds to the average of the corresponding values ​​of the patterns in this part. As a non-limiting alternative, the pattern values ​​used to calculate the average can be weighted by a quality factor associated with the corresponding pattern, for example, based on a similarity rate of this pattern relative to the other patterns in the signal.

[0100] Without limitation, this signal is selected here if the average patterns of this signal meet one or more similarity criteria. The similarity criteria may be based on a correlation and / or a difference in absolute values.

[0101] In this particular example, for each of the signals, step 15 includes a selection of that signal if: - the correlation coefficient of the average patterns of this signal is greater than a predetermined coefficient, for example greater than 0.95, and / or - the absolute difference between the mean patterns of this signal, following at least one pattern parameter, is less than a predetermined value, for example a systolic peak time difference of less than 7 ms.

[0102] The pattern parameters for the average patterns can be the same as those used for the initial patterns and can therefore be chosen from a list including, but not limited to, a peak time, a systolic peak time, a diastolic peak time, an interval between the systolic peak time and the diastolic peak time, a dicrotic notch time and an amplitude difference between a pattern start time and a pattern end time.

[0103] Datasets containing a signal selected at the end of operation 15 are called "processed datasets".

[0104] With reference to [Fig.1], the evaluation operation 9 includes a step of evaluating the quality of the processed datasets for each of the series, in order to determine whether the processed data of each of the individuals are usable for the training phase 2.

[0105] In this example, this step includes, for each of the processed datasets of each series, a calculation of a correlation between one or more parameters of the mean patterns of the signal of this set - the mean patterns being in this example calculated during step 15 of the processing 8 - and the value of systolic and / or diastolic blood pressure.

[0106] If the correlation thus calculated is greater than a predetermined coefficient, for example 0.9, for at least one of the parameters and for all the processed datasets of a given series, the datasets of this series are selected.

[0107] In this example, operation 9 also includes a step of evaluating the homogeneity of the set of series of datasets thus selected.

[0108] This step typically allows one or more of said series to be discarded and / or one or more new series to be acquired, so as to obtain a set of series of datasets which is homogeneous according to one or more homogeneity criteria.

[0109] For the sake of indication, said homogeneity criteria may refer to a homogeneous inter-series distribution in terms of blood pressure values, and / or in terms of age and / or gender of the corresponding individuals.

[0110] Phase 1, which has just been described, thus makes it possible to construct datasets forming several series, each corresponding to a respective individual.

[0111] In the non-limiting example of [Fig. 1], the datasets thus constructed are used to train the model, in the manner described below. Alternatively, the datasets resulting from the implementation of phase 1 can be used for other purposes, for example for analysis not implementing an artificial intelligence model.

[0112] With reference to [Fig. 1], phase 2 of the method is a model training phase using sets of training data formed from the sets of data constructed during phase 1 described above. Of course, the training data can be constructed differently, for example by processing collected data in a way that includes only some of the selection steps described above and / or one or more other processing steps.

[0113] In this example, training phase 2 includes a step in which: - one of the training datasets from one of the series is used to train so-called personalization data constituting a first input to the model, the personalization data comprising in this example, on the one hand, the average patterns of this set obtained during step 15 (see above) and, on the other hand, at least one of the blood pressure values ​​from this set, - one of the other training datasets in this series is used to train so-called test data, constituting a second input to the model; the test data in this example includes the average patterns of this dataset obtained during step 15 (see above), - an output of the model is formed by the difference between at least one of the blood pressure values ​​from the game forming the test data and at least one of the blood pressure values ​​from the personalization data.

[0114] This step is repeated iteratively using as test data each of the training datasets from the same series that do not form the personalization data, and iteratively using as personalization data each of the training datasets from the same series. These multiple combinations improve the learning.

[0115] This step and these iterations are then repeated in a similar manner for each of the other series.

[0116] Optionally, learning can be improved by using a third input formed by at least one parameter of the average patterns of the set forming the test data and / or of the set forming the personalization data.

[0117] The signals and patterns forming the inputs of the model can be one-dimensional, i.e. of a time nature, or two-dimensional, by carrying out for example a prior time-frequency transformation of the corresponding signals or patterns.

[0118] In a manner known per se, the training can be carried out using a conventional training algorithm, for example a backpropagation gradient algorithm.

[0119] Regarding phase 3 of the method in [Fig.1], this allows the input data of the model to be constituted during the prediction phase 4.

[0120] In general, phase 3 includes, for at least one individual who is typically different from those whose data were used to train the model, here referred to as the "target individual": - an operation 21 involving the collection of one or more datasets intended to form so-called personalization datasets, and - an operation 22 of collecting one or more signals intended to form so-called prediction signals.

[0121] In this example, for each target individual, operation 21 comprises a consecutive acquisition of three datasets analogous to the datasets of a grouping obtained during collection operation 6 of phase 1, the corresponding description applying by analogy.

[0122] Thus, in this example, for each target individual, each of the personalization datasets will be built on the basis of a set comprising: - a cardiac signal comprising patterns that each correspond to a respective heartbeat of the corresponding individual, - a systolic blood pressure reading for this individual, - a diastolic blood pressure value for this individual.

[0123] The signals acquired during collection operation 22 are cardiac signals similar to the signals acquired during collection operation 6 of phase 1, the corresponding description applying by analogy.

[0124] Thus, in this example, for each target individual, the prediction signals are constructed on the basis of photoplethysmographic signals, each comprising patterns which each correspond to a respective heartbeat of that individual.

[0125] In the example in [Fig. 1], phase 3 of the method comprises a preprocessing operation 23 and a processing operation 24 of these different data, which are analogous to operations 7 and 8 of phase 1, the corresponding description of which applies by analogy, a step such as step 11 of phase 1 being implemented, of course, only for the personalization data. These operations 23 and 24 thus make it possible to construct said personalization datasets and said prediction signals. In this example, phase 3 of the method does not, of course, implement steps analogous to those of operation 9 described above.

[0126] Of course, in the event that the data collected from a given target individual are not usable, phase 3 can be repeated in whole or in part in order to obtain usable input data.

[0127] During prediction phase 4, the prediction data for a given target individual are used as input data for the model: - The personalization datasets built during phase 3 constitute a first input to the model; the personalization data in this example includes, on the one hand, average patterns from this dataset obtained during step 24 and, on the other hand, at least one of the blood pressure values ​​from this dataset. - the prediction signals each iteratively constitute a second input to the model, the prediction signals in this example including average patterns obtained during step 24.

[0128] During prediction phase 4, the model provides as output, for each target individual, a blood pressure value for each of the prediction signals.

[0129] In this example, for each target individual, processing of the blood pressure values ​​provided by the model can be carried out by, for example, removing outliers and calculating the average of the other values.

[0130] Fig. 3 schematically represents a device 30 enabling the implementation of the method of Fig. 1.

[0131] The device 30 comprises a processing unit 31 and two sensors 32 and 33.

[0132] By way of non-limitation, the processing unit 31 is configured to implement each of phases 1 to 4 of the method in [Fig.1].

[0133] The processing unit 31 can generally form a hardware module such as a processor, an electronic chip, a calculator, a computer or a server, and a software module such as an application, a computer program or a virtual machine.

[0134] The processing unit 31 can be integrated into a device such as a smartphone, a tablet or a computer.

[0135] In this example, sensor 32 is a contact photoplethysmography sensor configured to measure a heartbeat signal by being placed in contact with an individual's skin and sensor 33 is a sphygmomanometer for measuring an individual's blood pressure.

[0136] Numerous variations can be made to the preceding description without departing from the scope of the invention. In particular, each of the phases of the method in [Fig. 1] can be implemented without resorting to one or more of the other phases of this method, or by replacing one or more of these phases with techniques different from those described above.

[0137] Among other embodiments of the invention, the signals used may be signals acquired using a non-contact photoplethysmography (rPPG) technique or a different technique, for example, a pulse signal acquisition technique using a pressure sensor. More generally, the data may be periodic or quasi-periodic in nature.

[0138] In an alternative embodiment, not shown, the construction of the data that is the subject of phase 1 of the method in [Fig. 1] is carried out in a process lacking a collection operation, the data to be processed in this case being available in a database. Of course, such a process may also lack one or more other steps and / or operations that are the subject of phase 1 of the method in [Fig. 1]. By way of example, the invention thus covers a process comprising one or more of the steps described above with reference to [Fig. 2] and not comprising operations 6, 7 and 9 of phase 1 of the method in [Fig. 1].

Claims

Demands

1. A method for processing (1) data including at least one signal which forms a plurality of patterns, for example a photoplethysmographic signal whose patterns are representative of an individual's heartbeats, the method comprising: - a measurement of at least one signal using at least one sensor, - a comparison of several of said patterns with respect to each other, - a selection of patterns which respect at least one predetermined similarity criterion, - a construction of average patterns each from at least two respective of said selected patterns, - a comparison of the average patterns with respect to each other, - a selection of the average patterns if these respect at least one predetermined similarity criterion.

2. Processing method (1) according to claim 1, wherein at least one similarity criterion is chosen from a list including: - a correlation coefficient greater than a predetermined value, - a difference between a value of at least one pattern parameter and a predetermined value less than a predetermined difference, the at least one parameter being able to be chosen from a list including a peak time, a systolic peak time, a diastolic peak time, an interval between the systolic peak time and the diastolic peak time and a dicrotic notch time.

3. A processing method (1) according to claim 1 or 2, wherein said plurality of patterns of at least one signal are patterns selected from initial patterns meeting at least one predetermined quality criterion, the at least one quality criterion being selectable from a list including:

4.

5. - a signal-to-noise ratio lower than a predetermined value, - a difference between a value of at least one pattern parameter and a predetermined value less than a predetermined difference, the at least one parameter being able to be chosen from a list including a peak time, a systolic peak time, a diastolic peak time, an interval between the systolic peak time and the diastolic peak time, a dicrotic notch time, an instantaneous heart rate, an instantaneous heart rate variation, a pattern amplitude and an amplitude difference between a pattern start time and a pattern end time. Processing method (1) according to any one of claims 1 to 3, wherein said at least one signal comprises several signals and wherein said physiological data comprise one or more groups each comprising several data sets, each of these data sets comprising one respective of said signals as well as at least one blood pressure value, the method comprising a selection of the group or groups in each of which the blood pressure values ​​of the data sets of that group are within a corresponding predetermined range of values ​​and / or exhibit variability less than a predetermined value. A processing method (1) according to any one of claims 1 to 4, comprising a data collection operation (6), the data collection operation (6) preferably comprising, for each individual among one or more individuals, the acquisition of several groups of datasets, each dataset comprising: - a signal that forms a plurality of patterns, preferably a photoplethysmographic signal whose patterns are representative of the heartbeats of that individual, - at least one blood pressure reading for this individual, the signals being acquired consecutively within each of the groups.

6. A method (2) for training an artificial intelligence model using training data formed from data processed by a processing method (1) according to any one of claims 1 to 5, the training data preferably being formed from data collected according to a method having the characteristics of claim 5 and used such that: - a first input of the model receives data obtained on the basis of the signal and at least one blood pressure value from one of the datasets of one of the individuals, this dataset being called the personalization set, - a second input of the model receives data obtained on the basis of the signal from another of said datasets of this individual, this other set being called the test set,- an output of the model is formed by the difference between at least one blood pressure value from the test set and at least one blood pressure value from the customization set.

7. Method for estimating (4) a physiological parameter such as a blood pressure value of one or more individuals, using an artificial intelligence model trained with a training method (2) according to claim 6.

8. Device (30) comprising a processing unit (31) configured to implement a method according to any one of claims 1 to 7 and at least one sensor (32, 33) for measuring at least one physiological signal and / or at least one physiological parameter.

9. Computer program comprising executable computer instructions which, when executed by computer, implement a method according to any one of claims 1 to 7.

Images