Method for determining a score indicating the risk of an individual having an epileptic seizure, method for assisting with detection and / or prediction of a seizure implementing such a method, and corresponding systems

Abstract

The invention relates to a method for determining a seizure risk score and to a method for predicting and / or detecting epileptic seizures of a patient, said methods comprising: a step of building (110) a learning database of epileptic individuals; a step of partitioning (120) said learning database into a plurality of clusters; for each cluster, a step of training (130) a machine-learning model specific to this cluster, using meta-indicators derived from the data of the learning database restricted to individuals belonging to said cluster; a step of classifying (140) the patient into a cluster selected from said plurality of clusters, so as to assign to the patient the corresponding trained learning model; a step of acquiring (150) ECG and accelerometer signals from the patient, who is fitted with an ECG sensor and an accelerometer; a step of computing the same meta-indicators as those of the learning phase; a step of computing a score (160) indicating the risk of the patient having a seizure, by feeding the trained learning model corresponding to the cluster of the patient with the computed meta-indicators.

Description

[0001] DESCRIPTION

[0002] TITLE OF THE INVENTION: METHOD FOR DETERMINING AN EPILEPTIC SEIZURE RISK SCORE OF AN INDIVIDUAL, METHOD FOR ASSISTANCE IN THE DETECTION AND / OR PREDICTION OF SEIZURES IMPLEMENTING SUCH A METHOD, AND CORRESPONDING SYSTEMS

[0003] Technical field of the invention

[0004] The invention relates to a system and method for assisting in the detection and / or prediction of epileptic seizures in a patient.

[0005] Technological background

[0006] Epilepsy is a disabling neurological disease that affects tens of millions of people worldwide. Despite therapeutic advances, approximately one-third of these patients suffer from treatment-resistant forms of epilepsy and still regularly experience seizures. According to a definition proposed by the International League Against Epilepsy, these drug-resistant forms occur after the failure of two tolerated and correctly chosen antiepileptic treatments (alone or in combination) to achieve lasting seizure control. Unbalanced patients constitute the most worrying population, for whom a mortality rate is observed that is significantly higher than that of the rest of the population.In France, several hundred thousand people are affected by epilepsy and the number of drug-resistant patients is estimated at more than 150,000, which represents a significant medical and social cost for the health system.

[0007] Epilepsy is a complex and multifaceted disease characterized by variable etiologies, ages of onset, and clinical manifestations. The International League Against Epilepsy proposes a classification of seizures based on the location of the initial manifestations (focal or generalized), the level of consciousness, motor signs, and the possible spread of the seizure. All seizures can be prolonged and lead to status epilepticus, which represents a major medical emergency. Epilepsy may also be accompanied by cognitive or psychiatric comorbidities.

[0008] In the specific case of focal epilepsy, surgical treatment may be considered, although in practice it is only possible in a limited proportion of patients. In unbalanced and inoperable patients, the unpredictable nature of seizures is a significant source of anxiety. The quality of life of these people and their loved ones is consequently significantly reduced.

[0009] Seizure frequency measurement is considered an essential indicator for assessing the quality of care. Several studies have highlighted the difficulty patients face in maintaining a reliable seizure diary. Phenomena such as memory loss or altered perception at the time of a seizure are thought to be responsible. Whatever the reasons, it has been observed that at least 50% of seizures are missed on average by patients. A study of patients implanted for several months with an intracranial device (allowing very precise seizure detection) also confirmed that most underestimated the frequency of their seizures. These findings motivate the development of robust outpatient seizure detection systems.

[0010] There are already commercially available systems for detecting certain types of seizures that present significant motor manifestations such as generalized tonic-clonic seizures. However, these events represent a minority of seizures.

[0011] More broadly, various detection approaches have been tested. They are generally based on the identification of biomarkers calculated on one or more physiological parameter(s). The biological signals monitored are as diverse as electroencephalogram (better known by the acronym EEG), video, audio, movements, heart rate, skin conductance, skin temperature, oxygenation rate, respiration or even muscular activity via an electromyogram (or EMG). Multimodal approaches, i.e. combining several sensors at the same time, are considered the most promising (better sensitivity and specificity).

[0012] However, there is currently no non-invasive system that can reliably detect all epileptic seizures, particularly focal seizures.

[0013] The inventors therefore sought to develop a new system and method that makes it possible to detect and / or anticipate epileptic seizures in order to improve the monitoring and comfort of patients and their loved ones.

[0014] Objectives of the invention

[0015] The invention thus aims to provide a system and a method for determining a patient's epileptic seizure risk score.

[0016] The invention also aims to provide a system and a method for assisting in the detection and / or prediction of epileptic seizures in a patient implementing the system and method for determining a seizure risk score according to the invention.

[0017] The invention also aims to provide, in at least one embodiment, such a system and such a method which allow the detection and / or prediction of epileptic seizures, regardless of the symptoms announcing the seizures.

[0018] The invention also aims to provide, in at least one embodiment, such a system and such a method which make it possible to limit the risks of accidents linked to seizures, to reassure patients and their loved ones on a daily basis and finally to improve the quality of care by a neurologist.

[0019] The invention also aims to provide, in at least one embodiment, such a system and such a method which can be used by a variety of individuals, without particular constraints.

[0020] The invention also aims to provide, in at least one embodiment, such a system and such a method which can evolve over time based on user feedback.

[0021] Statement of the invention

[0022] To do this, the invention relates to a method for determining an epileptic seizure risk score of an individual, called a patient, said method comprising: a step of constituting a learning base of epileptic individuals, said base comprising, for at least a plurality of individuals, electrocardiogram signals, called ECG signals, and accelerometric signals, respectively representative of the electrical activity of the heart and the movements of these individuals, associated with data on the start and end of seizures of these individuals, a step of partitioning said learning base into a plurality of groups of individuals, called clusters, each cluster being made up of individuals presenting the same type of epileptic symptom(s) in the vicinity of the epileptic seizures that they undergo.

[0023] Furthermore, the method comprises, for each cluster formed in the previous step, a step of training a machine learning model specific to this cluster, from meta-indicators derived from the data of said learning base restricted to the individuals belonging to this cluster, said meta-indicators being calculated by the following steps: localization of all the R peaks in the ECG signal of each individual of the cluster, an R peak being the most prominent peak of a QRS complex generated by a heartbeat of the individual (contraction of the ventricles), identification of the RR intervals of the ECG signal of each individual of the group, calculation of indicators, called cardiac indicators, from the RR intervals including a mean, a standard deviation, a low and high frequency power, the cardiac indicators being calculated at a predetermined time interval on the ECG signals of each individual, calculation of indicators, called activity indicators,from the accelerometric signals of each individual in the cluster, the activity indicators being calculated at a predetermined time interval on the accelerometric signals of each individual, operations on said indicators calculated in the previous steps to provide said meta-indicators representative of a distribution of the indicators within a predetermined time window, of each cardiac and activity indicator calculated.,

[0024] The method also comprises the following steps: a step of classifying said patient into one of the clusters so as to select the corresponding trained learning model, a step of acquiring ECG and accelerometric signals on said patient equipped with an ECG sensor and an accelerometer, a step of calculating said meta-indicators following the same principle as that implemented for training the learning models, a step of calculating a crisis risk score for said patient, by feeding said trained learning model corresponding to the group of said patient with said meta-indicators calculated for this patient.

[0025] The method according to the invention therefore comprises a first series of steps (hereinafter referred to as the pre-learning phase) which makes it possible to constitute a learning base, to partition this learning base into a plurality of groups (referred to as clusters throughout the following), and to train a learning model for each cluster. This pre-learning phase therefore makes it possible to obtain a plurality of trained learning models, each model corresponding to a typology of individuals grouped by type of epileptic symptom(s).

[0026] The method further comprises a second series of steps (hereinafter referred to as the evaluation phase) which makes it possible to equip a patient with a plurality of sensors, to calculate different meta-indicators from the sensor data and to feed the trained learning model corresponding to the patient's group membership with the calculated meta-indicators to obtain a seizure risk score. This evaluation phase occurs after the pre-learning phase and makes it possible to determine a risk score for a patient equipped with an ECG sensor and an accelerometer. Once the pre-learning phase has been carried out, the patient evaluation phase can, of course, be executed several times, without the need to repeat the pre-learning phase.

[0027] It has been observed in some patients that the onset of a crisis is concomitant with a modification of the autonomic nervous system, which can be observed in the variability of the heart rate.

[0028] These changes are not the same in all patients (for example, some have tachycardia, others bradycardia).

[0029] Similarly, in some patients, seizures are accompanied by gestures (or the absence of gestures) that can be captured by acceleration sensors.

[0030] Given this diversity of manifestations, the method according to the invention trains a machine learning model specific to each type of manifestation to improve the quality of detection. Thus, rather than having a model specific to all the individuals in the training base constituted, the invention provides for assigning the patient to a group of individuals from among a plurality of groups and training and using a machine learning model per group thus constituted.

[0031] It was therefore planned to learn a model for each group of patients who exhibited the same types of symptoms around the time of the attacks. The automatic learning of the model for each group (cluster) is done in the pre-learning phase.

[0032] The method according to the invention comprises a step of partitioning the learning base into a plurality of groups of individuals, called clusters, each cluster being made up of individuals presenting the same type of epileptic symptom(s) in the vicinity of the epileptic seizures that they suffer.

[0033] According to an advantageous variant of the invention, this partitioning step comprises: a step of randomly distributing said individuals from said learning base into K clusters, K being a predetermined integer, for each of the K clusters thus formed, a step of training a machine learning model specific to this cluster according to the same principle as that implemented in the context of training learning models from said meta-indicators derived from the data of the individuals in this cluster, a reassignment of the individuals from said learning base to the cluster whose learning model minimizes a prediction error, a reiteration of the preceding steps until said total prediction error stabilizes around a value.

[0034] In this variant, a number K of clusters is arbitrarily set and these K clusters are randomly formed. For each cluster thus formed, a classification model is trained following the same principle as the training of learning models described later. Once all the models are trained, the individuals are reassigned to the cluster whose model minimizes the prediction error. This process is repeated several times until the total prediction error stabilizes around a value.

[0035] This procedure advantageously makes it possible to constitute clusters of individuals without a priori knowledge of the individuals' symptoms.

[0036] In another variant, clusters are formed by an operator that analyzes the learning base and groups individuals according to the individuals' seizure symptoms.

[0037] The method according to the invention also comprises, for each cluster formed in the previous step, a step of training a machine learning model specific to this cluster, from meta-indicators derived from the data of the learning base restricted to the individuals belonging to this cluster.

[0038] The meta-indicators are calculated by the following steps: localization of all the R peaks in the ECG signal of each individual in the cluster, an R peak being the most prominent peak of a QRS complex generated by a heartbeat of the individual, identification of the RR intervals of the ECG signal of each individual in the group, calculation of indicators, called cardiac indicators, from the RR intervals including a mean, a standard deviation, a low and high frequency power, the cardiac indicators being calculated at a predetermined time interval on the ECG signals of each individual, calculation of indicators, called activity indicators, from the accelerometric signals of each individual in the cluster, the activity indicators being calculated at a predetermined time interval on the accelerometric signals of each individual,operations on said indicators calculated in the previous steps to provide said meta-indicators representative of a distribution of the indicators within a predetermined time window, of each cardiac and activity indicator calculated.,

[0039] The meta-indicators thus calculated are the input data provided to the learning models implemented by the invention.

[0040] If other sensors are used to enrich the learning base with other signals, other meta-indicators can also be calculated to feed the learning models.

[0041] Once all the meta-indicators of all the individuals in the training base are calculated, the machine learning models can be trained. According to the invention, a machine learning model is trained by groups of individuals formed in the partitioning step.

[0042] This step closes the pre-learning phase of the method according to the invention and the patient evaluation phase can then be implemented.

[0043] The evaluation phase aims to provide a risk score for a patient equipped with an ECG sensor and an accelerometric sensor from which the same meta-indicators as those calculated during the pre-training phase will be determined and will feed the trained learning model corresponding to the cluster to which the patient belongs.

[0044] To do this, the patient is, for example, equipped with a box that houses several sensors, including an ECG sensor and an accelerometer. These sensors communicate, for example via Bluetooth, with an application embedded on a device such as a smartphone. This embedded application is configured to calculate the meta-indicators from the data provided by the sensors. In addition, the software application is configured to retrieve the trained learning model corresponding to the patient. Thus, the meta-indicators can feed the trained learning model corresponding to the patient to provide a seizure risk score. Note that it is also possible to directly integrate an electronic card configured to allow the calculation of the meta-indicators directly within the box or to program the sensors so that the meta-indicators are directly calculated by them.

[0045] Thus, this evaluation phase begins with a step of classifying the patient into one of the clusters in order to select the corresponding trained learning model.

[0046] Advantageously and according to the invention, said step of classifying said patient in a cluster comprises: a step of acquiring ECG and accelerometric signals on said patient from an ECG sensor and an accelerometer worn by said patient, a step of calculating said meta-indicators following the same principle as that implemented for training the learning models, a step of reporting by said patient the first N seizures suffered, N being a predetermined integer, a step of calculating the risk score from each of the trained learning models, and of selecting said trained model which provides the best risk score for all of the first N seizures reported, said patient being classified in the cluster corresponding to this model.

[0047] According to this advantageous variant, the step of classifying the patient within a cluster is done automatically by testing the different learning models. This makes it possible to select a pre-learned model that best corresponds to the first N seizures reported by the patient. Once a cluster has been assigned to the patient, the method according to the invention can calculate the patient's seizure risk score, by feeding the trained learning model corresponding to the patient's group with the meta-indicators calculated for this patient.

[0048] Advantageously and according to the invention, the method further comprises a step of indicating a crisis prediction error by said patient and a step of updating the learning model of said patient from said learning base restricted to individuals belonging to the cluster of said patient, to which said patient is added.

[0049] In this advantageous variant, the trained learning model corresponding to the patient's cluster is updated if the patient (or a relative of the patient, a practitioner, etc.) reports prediction errors. This makes it possible to improve the learning model used by the patient. The prediction of a seizure is defined, for example, as a risk score above a predetermined threshold. Also, a prediction error corresponds to a calculated score above the predetermined threshold while the patient (or a relative, a practitioner, etc.) believes that no seizure has occurred.

[0050] To do this, the pre-learning phase limited to the patient's learning model is carried out again with the patient's data corresponding to the prediction errors indicated by the patient (or by a relative, a practitioner, etc.). These data are formed from the meta-indicators corresponding to the prediction errors indicated by the patient (or by a relative, a practitioner, etc.).

[0051] If several patients are equipped with a system according to the invention, all the data from the different patients can usefully enrich the learning base of the invention, provided of course that these patients belong to the same cluster.

[0052] The invention also relates to a method for assisting in the detection and / or prediction of epileptic seizures of an individual, called a patient, said method comprising: a method for determining an epileptic seizure risk score for said patient according to the invention, a step of detecting and / or predicting a possible epileptic seizure by comparing said seizure risk score determined by said determination method with at least one predetermined threshold.

[0053] The advantages and technical effects of the method for determining a risk score according to the invention apply mutatis mutandis to the method for assisting in the detection and / or prediction of crises according to the invention.

[0054] Furthermore, the method according to the invention makes it possible to detect a possible crisis or a prediction of a crisis by comparing the risk score determined by the invention with one or more predetermined thresholds. It is for example possible to monitor the evolution of the risk score and to deduce a risk of crises by detecting an evolution of the risk score over a predetermined period of time.

[0055] The invention can also be usefully associated with a patient seizure diary system allowing the patient and his neurologist to better calibrate treatments and to objectively and reliably observe the evolution of the disease.

[0056] The invention may also be usefully associated with a system for monitoring the duration and severity of the patient's seizures.

[0057] The invention can also be usefully associated with a patient alert system in the event of detection of an imminent crisis so that the patient can get to safety.

[0058] The invention can also be usefully associated with a system for alerting relatives of the patient or caregivers, in association with means of geolocating the patient to enable the patient to be found and secured.

[0059] The invention can also be usefully associated with means for automatically recording raw signals (ECG and accelerometer) in the event of seizures in order to allow the neurologist to review the tracings during a consultation, for example. The invention can also be usefully associated with a system for anticipating the risk of seizures based on the recurrence of past seizures observed in a patient.

[0060] The invention also relates to a system for determining a seizure risk score for an individual, called a patient, said system comprising: a learning base for epileptic individuals, said base comprising, for at least a plurality of individuals, electrocardiogram signals, called ECG signals, and accelerometric signals, respectively representative of the heartbeats and movements of these individuals, associated with data on the start and end of seizures of these individuals, a module for partitioning said learning base into a plurality of groups of individuals, called clusters, each cluster being made up of individuals presenting the same type of epileptic symptom in the vicinity of the epileptic seizures that they undergo, for each cluster formed in the previous step, a machine learning model specific to this cluster,trained from meta-indicators derived from the data of said learning base restricted to the individuals belonging to this cluster, said meta-indicators being provided by a processing unit comprising: o a module for locating all the R peaks in the ECG signal of each individual of the cluster, an R peak being the most prominent peak of a QRS complex generated by a heartbeat of the individual, o a module for identifying the RR intervals of the ECG signal of each individual of the cluster, o a module for calculating indicators, called cardiac indicators, from the RR intervals including a mean, a standard deviation, a low and high frequency power, the cardiac indicators being calculated at a predetermined time interval on the ECG signals of each individual, o a module for calculating indicators, called activity indicators, from the accelerometric signals of each individual of the cluster,the activity indicators being calculated at a predetermined time interval on the accelerometric signals of each individual, o a module for combining said calculated indicators to provide said meta-indicators representative of a distribution of the indicators within a predetermined time window, for each calculated cardiac and activity indicator, a module for classifying said patient in a cluster chosen from said plurality of clusters defined by the partitioning step, so as to select the corresponding trained learning model, a box worn by said patient comprising at least one ECG sensor and an accelerometer, a module for calculating said meta-indicators following the same principle as that implemented for training the learning models, a module for calculating a crisis risk score for said patient,by feeding said learning model corresponding to the cluster of said patient with said meta-indicators calculated for this patient.,

[0061] The advantages and technical effects of the method according to the invention apply mutatis mutandis to a system according to the invention.

[0062] Furthermore, a system according to the invention advantageously implements a method according to the invention and a method according to the invention is advantageously implemented by a system according to the invention.

[0063] Throughout the text, a module is defined as a software element, a subset of a software program, which can be compiled separately, either for independent use or to be assembled with other modules of a program, or a hardware element, or a combination of a hardware element and a software subroutine. Such a hardware element may include an application-specific integrated circuit (better known by the acronym ASIC for the English term Application-Specific Integrated Circuit) or a programmable logic circuit (better known by the acronym FPGA for the English term Field-Programmable Gate Array) or a specialized microprocessor circuit (better known by the acronym DSP for the English term Digital Signal Processor) or any equivalent hardware or any combination of the aforementioned hardware. Generally speaking, a module is therefore an element (software and / or hardware) which makes it possible to perform a function.

[0064] The system according to the invention has the particularity of combining data derived from an electrocardiogram and an accelerometer worn by an individual to determine a risk of seizure from a machine learning model trained for this purpose. The machine learning model implemented is furthermore dependent on a group of individuals to which the monitored patient is attached. The system according to the invention thus implements a machine learning model per group of individuals.

[0065] The principle of use of the system according to the invention is as follows.

[0066] When a patient equips himself with the system according to the invention, a rapid pre-learning phase can be implemented to assign him a specific cluster. This step aims, from a reduced data set, to determine the group to which the patient belongs. This data is for example data from previous seizures reported by the patient himself, which makes it possible to select the typology of symptom(s) that he presents and thus the most appropriate model to use. The model thus selected makes it possible to initialize a model specific to the patient. The system can operate in continuous mode. The patient's model is not fixed: as soon as the latter gives feedback on a detection error or a missed seizure, the system can change the group to which he belongs and / or update the trained model.

[0067] The data provided by the sensors are collected in real time, and the same indicators as those used in the pre-learning phase are calculated by the different modules of the system.

[0068] The patient may also be asked to report the first N attacks suffered via an application or a personal space on a website, where N is a predetermined integer. This allows a pre-learned model that best suits the patient to be selected.

[0069] At any time the patient can report a prediction error. If so, their model is updated.

[0070] To avoid over-learning, updating a patient model is done by including it in the patient cluster of the pre-learning database that was used to pre-learn the model.

[0071] The patient's trained learning model can be saved to an app embedded on the patient's smartphone. The latter contains the most up-to-date patient's trained learning model (its parameters are therefore fixed in the app).

[0072] Regularly (for example once a month), the server updates the parameters of its learning model based on patient feedback and sends them to the application. It is also possible to repeat the pre-learning phase offline with new data that the patient provides (or that different patients equipped with a system according to the invention provide).

[0073] The invention also relates to a system for assisting in the detection and / or prediction of epileptic seizures in an individual, called a patient, comprising: o a system for determining a seizure risk score according to the invention, o a module for detecting and / or predicting an epileptic seizure by comparing the risk score determined by said system for determining a seizure risk score with at least one predetermined threshold.

[0074] The advantages and technical effects of the device for determining a patient's seizure risk score apply mutatis mutandis to a system according to the invention. The invention also relates to a method for predicting and / or detecting seizures and a system for predicting and / or detecting seizures characterized in combination by all or some of the features mentioned above or below.

[0075] List of figures

[0076] Other aims, characteristics and advantages of the invention will appear on reading the following description given solely for non-limiting purposes and which refers to the appended figures in which:

[0077] [FIG 1] is a schematic view of the operating principle of a system for determining an epileptic seizure risk score according to one embodiment of the invention,

[0078] [FIG 2] is a synoptic view of the method for determining an epileptic seizure risk score according to one embodiment of the invention.

[0079] Detailed description of an embodiment of the invention

[0080] Figure 1 schematically illustrates the principle of determining a patient's epileptic seizure risk score using the method schematically represented by Figure 2.

[0081] The method according to the invention comprises two distinct phases: a pre-learning phase and a phase of evaluating a patient using the results of the pre-learning phase.

[0082] The pre-learning phase comprises a first step 110 of constituting a learning base 20. The method comprises a subsequent step 120 of partitioning the learning base 20 into a plurality of clusters 20a, 20b, 20c, 20d. The method then comprises a step 130 of training a learning model per cluster. At the end of this step, we have models 40a, 40b, 40c, 40d corresponding to the clusters 20a, 20b, 20c, 20d.

[0083] The evaluation phase comprises a step 140 of classifying the patient 10 in a cluster so as to select the relevant trained learning model for the evaluation of the patient 10. To do this, and according to one embodiment, cardiac and activity signals are recovered from the patient and compared to the first N attacks reported by the patient (these reports being for example made via a software application embedded on the patient's smartphone 12).

[0084] The data of the selected model can then be transmitted in a step 142 to a mobile application embedded on the patient's smartphone 12. According to another variant, all of the models are transmitted to the patient's smartphone and the classification phase is directly implemented on the patient's smartphone.

[0085] The evaluation phase then comprises a step 150 of acquiring ECG and accelerometric signals on the patient from the box 14 worn by the patient, which then makes it possible to carry out a step 160 of calculating a risk score by submitting this data, from which meta-indicators will be extracted, to the trained learning model selected for the patient. The model provides the corresponding risk score, a comparison 170 of which with one or more predetermined thresholds makes it possible to signal a potential epileptic seizure.

[0086] The different stages and the means that can be used to implement them will now be described in more detail.

[0087] Pre-learning phase

[0088] The pre-learning phase begins with the creation of a learning base 20.

[0089] Learning base

[0090] To constitute the learning base, it is for example possible to equip a plurality of epileptic individuals 18 with a box according to the invention and to continuously monitor the individuals 18 thus equipped.

[0091] This box houses at least one ECG sensor and one accelerometer configured to acquire corresponding signals from individuals. This box can be of any type and can be worn by each individual, either as a patch, a necklace, a watch, etc. This box is for example formed by a belt-type device known under the name BioHamess 3.0 from the Zephyr® brand (Medtronic®). This device allows continuous collection of an electrocardiogram (ECG) (sampled at 1 kHz) and 3D accelerometry data (sampled at 100 Hz).

[0092] It should also be noted that this equipment also makes it possible to collect data on the deformation of the rib cage, allowing breathing to be reported (sampled at 25 Hz). As described later, these data can, according to one embodiment, enrich the learning base of the invention.

[0093] This box can be used both to constitute the learning base, but also to equip the patient for whom we then seek to determine a risk score.

[0094] The applicant then carried out data measurements for an average of 17 days (maximum 21 days, minimum 12).

[0095] During this period, an average of 23 seizures per individual were detected, a maximum of 107 and a minimum of 2. The signals from the ECG and accelerometer sensors were recorded continuously and the onset and end of seizures were notified by the caregivers who monitored the individuals. All of this data constitutes the learning base of the invention. It is also possible, to constitute the learning base and associate the acquired signals with the information on the start and end of seizures, to monitor individuals with other types of equipment, such as intracranial devices.

[0096] It is also possible, to enrich the learning base thus constituted, to use an already constituted database 19, such as for example the European Epilepsiae database.

[0097] The applicant used the Epilepsiae database, which consists of recordings made in 3 European centers (La Pitié Salpétrière in France, Lreiburg in Germany, and Coimbra in Portugal). Data from 168 individuals (out of the 275 individuals that make up the complete database) were retrieved by the applicant to enrich his learning database. The Epilepsiae database includes several days of data per individual (8 on average, 24 maximum, 3 minimum).

[0098] The following data were used for each selected individual: ECG data, age, date of admission, neurological assessment, treatment, assessments, seizure frequency, seizure type,

[0099] Metadata for each seizure: seizure onset and end on EEG, seizure onset and end from a clinical perspective (video), seizure type.

[0100] Note that this Epilepsiae database does not include accelerometric data from individuals. This portion of the database therefore does not allow the corresponding meta-indicators to be obtained. However, the learning models of the method according to the invention can be fed with missing data. At least some of the individuals in the database include ECG signals and accelerometric signals, as indicated previously.

[0101] According to one embodiment of the invention, the learning base 20 can further be enriched with other data, such as for example the body temperature of the individuals and / or the electrodermal activity of the individuals.

[0102] Partitioning the learning base

[0103] Once this learning base 20 has been constituted, it is partitioned into a plurality of clusters 20a, 20b, 20c, 20d by a partitioning module.

[0104] This step 120 of partitioning the base into a plurality of clusters 20a, 20b, 20c, 20d is preferably carried out automatically.

[0105] To do this, the individuals from base 20 are initially randomly distributed within clusters 20a, 20b, 20c, 20d.

[0106] For each of the clusters thus formed, a machine learning model specific to this cluster is trained from the meta-indicators derived from the data of the individuals in this cluster (described later).

[0107] Then, a reassignment of individuals from the learning base to the cluster whose learning model minimizes a prediction error is performed.

[0108] These steps are repeated until the prediction error stabilizes around a value.

[0109] The principle of partitioning according to this embodiment is similar to the implementation of a k-means type algorithm.

[0110] According to another embodiment, the clusters are formed by an operator who analyzes the learning base and groups the individuals according to the individuals' seizure symptoms.

[0111] Training learning models

[0112] Once the training base is partitioned, a learning model is trained for each cluster from meta-indicators derived from the data of that cluster. The meta-indicators are determined by the following sub-steps.

[0113] Peak localization and RR interval calculation

[0114] The first substep concerns the localization 131 of the R peaks in the ECG signal. This is the most prominent peak in the characteristic wave called QRS complex that a heartbeat generates. It is then possible to determine the RR intervals of the signal in a second substep 132 of identifying the RR intervals. The scientific literature describes routines for determining RR intervals from ECG signals and the person skilled in the art can refer to it to implement this step of the method.

[0115] Calculation of cardiac indicators

[0116] Once the different RR intervals have been detected in the signal, it is possible to calculate a certain number of cardiac indicators during a cardiac indicator calculation sub-step 133.

[0117] The following indicators can be used: mean, standard deviation, 12-norm of the derivative (these indicators can for example be obtained by following the principles outlined in the document “Malik, M., JT Bigger, AJ Camm, RE Kleiger, A. Malliani, AJ Moss, and PJ Schwartz. 1996. “Heart Rate Variability: Standards of Measurement, Physiological Interpretation, and Clinical Use.” European Heart Journal 17 (3): 354-81. https:

[0118] / / doi.org / 10.1093 / oxfordjournals.eurheartj.a014868 ». low and high frequency power (these indicators can for example be obtained by following the principles outlined in the aforementioned Malik et.al. document) CSI index - English acronym for Cardiac Sympathetic Index - (this indicator can for example be obtained by following the principles outlined in the document "Jeppesen, Jesper, Anders Euglsang-Erederiksen, Peter Johansen, Jakob Christensen, Stephan Wüstenhagen, Hatice Tankisi, Erisela Qerama, Alexander Hess, and Sândor Beniczky. 2019. "Seizure Detection Based on Heart Rate Variability Using a Wearable Electrocardiography Device". Epilepsia 60 (10): 2105-13. https: / / doi.org / 10.1111 / epi.16343.) entropy indicators (these indicators can for example be obtained by following the principles outlined in the document "Thu, Trang Nguyen Phuc, Alfredo I. Hernandez, Nathalie Costet, Hugues Patural, Vincent Pichot, Guy Carrault, and Alain Beuchée. 2019.“Improving Methodology in Heart Rate Variability Analysis for the Premature Infants: Impact of the Time Length”. PLOS ONE 14 (8): e0220692. https: / / doi.org / 10.1371 / journal. pone.0220692).

[0119] Calculation of activity indicators

[0120] The method according to the invention also calculates a certain number of indicators representative of the activity of individuals during a sub-step 134 of calculating the activity indicators.

[0121] The following indicators can be used: norms of the accelerometric signal on each of the 3 axes (x, y, z) of the accelerometer obtained by the following equations:

[0122] ||d|| y af af af represent the values acceleration measured on each of the three axes of a direct trihedron (x,y,z) associated with the accelerometer, standard of the complete signal defined by the following equation:

[0123] Other indicators can also be calculated from acceleration measurements.

[0124] Operations on indicators

[0125] Heart rate and activity metrics are calculated, for example, over sliding windows whose size can vary, which allows their number to be multiplied. For example, the method can calculate an indicator value every 10 seconds.

[0126] The method according to the invention then combines these different indicators to form meta-indicators during a sub-step 135 of calculating the meta-indicators. For example, every 10 seconds, the method retrieves all the indicator values ​​of the last 10 minutes and calculates meta-indicators for each indicator (average, variance, minimum, maximum, etc.)

[0127] It is these meta-indicators, representative of the distribution of indicators over the time window considered (for example the last 10 minutes) which feed the supervised learning model associated with each cluster.

[0128] An a priori crisis probability indicator can also be added to these meta-indicators. This probability indicator can be calculated from the individuals' crisis history and any cycles observed in individuals.

[0129] Alternatively, it is possible to add metadata corresponding to the patient (epileptic syndrome, etc.).

[0130] Learning model

[0131] Once all the meta-indicators of all the individuals in the learning base are calculated, the automatic learning models 40a, 40b, 40c, 40d can be trained during the training step 130. According to the invention, a machine learning model is trained per group of individuals formed in the partitioning step.

[0132] In one variant, support vector machines (SVMs) are used as training models. In another variant, ensembles of decision trees (Random Forest, XGBoost, CatBoost, or LightGBM) are used as training models. In another variant, deep neural network models are used as training models.

[0133] At the end of this step, trained models 40a, 40b, 40c, 40d are available for the different clusters 20a, 20b, 20c, 20d. All of the steps relating to the pre-training phase are implemented on a computer system, which must be understood in the broad sense (computer, plurality of computers, virtual server on the Internet, virtual server on the Cloud, virtual server on a platform, virtual server on a local infrastructure, server networks, etc.). This computer system typically comprises one or more processors, one or more memories and possibly a human-machine interface. The computer system also comprises a database for saving the learning database and the clusters obtained by partitioning the database. The computer system may also comprise means for storing the trained learning models.

[0134] Evaluation phase

[0135] Once the different learning models have been trained from the learning base, the patient assessment phase can be carried out. This assessment phase aims to calculate a seizure risk score and possibly detect and / or predict seizures in a patient equipped with an ECG sensor and an accelerometer.

[0136] To do this, a patient is equipped with a box 14 in which an accelerometer and an ECG sensor are housed. This box can, for example, be the same box as that used to constitute the learning base.

[0137] The accelerometer is preferably a 3-axis accelerometer. Alternatively, it is possible to use 3 single-axis accelerometers, or even one single-axis accelerometer.

[0138] These sensors are intended to provide ECG and accelerometric signals to a processing unit to calculate the same meta-indicators as those calculated during the pre-learning phase, these calculated meta-indicators then being passed into the trained learning model corresponding to the cluster to which the patient belongs.

[0139] This processing unit therefore comprises a module for calculating meta-indicators and a module for calculating a patient crisis risk score. As indicated previously, these modules may be software elements, subsets of a software program, hardware elements, or a combination of hardware elements and software subprograms. According to one variant, these modules are embedded on the patient's smartphone and use the smartphone's microprocessor and memories to perform the corresponding operations.

[0140] Thus and according to this embodiment, the sensors of the housing 14 communicate, for example via Bluetooth, with the application embedded on the patient's smartphone 12. This embedded application is configured to calculate the meta-indicators and to submit the trained learning model corresponding to the patient to the calculated meta-indicators. The application can thus communicate with a remote server to retrieve the corresponding learning model.

[0141] To determine the model corresponding to the patient, according to a preferred embodiment, the patient is first classified within a cluster during a classification step 140.

[0142] Patient classification within a cluster

[0143] Preferably, the classification of the patient within a cluster comprises a first step of acquiring 141 ECG and accelerometric signals on the patient from the box worn by the latter. Then, the meta-indicators are calculated according to the same principle as that implemented for training the learning models in a step of calculating 142 the meta-indicators. This classification step then preferably provides for a reporting 143 by the patient of the first N seizures suffered, N being a predetermined integer. This can for example be done by a dedicated interface on the mobile application embedded on the patient's smartphone. The risk score associated with each reporting is then calculated from each of the learning models trained during the learning phase.Finally, the trained model that provides the best risk score for all N first reported seizures is selected in a relevant model selection sub-step 144. The patient is classified into the cluster corresponding to this selected model.

[0144] Calculating the risk score

[0145] Once the training model is selected, the model data is downloaded to the patient's smartphone software application.

[0146] The system can then continuously acquire the sensor signals, calculate the meta-indicators and submit them to the learning model to obtain the risk score. This step of calculating a risk score 170 thus implements and preferably the same steps as those described in connection with the learning of the model dedicated to the patient. Updating the model

[0147] According to an advantageous embodiment, the patient can indicate, during a step 180, a crisis prediction error, for example via an interface of the software application.

[0148] In this case, it is possible to update the learning model used by the patient by retraining it following the same principle as described previously, but by integrating the data corresponding to the prediction error into the learning base. To avoid over-learning, the patient's data is then added to the learning base and a specific weight associated with this data can be provided to give more importance to the patient's feedback compared to the data of other individuals. This allows for a patient-specific learning model that will be increasingly relevant over time.

Claims

CLAIMS 1. Method for determining an epileptic seizure risk score of an individual, called a patient, said method comprising: a step of constituting (110) a learning base (20) of epileptic individuals, said base (20) comprising, for a plurality of individuals, electrocardiogram signals, called ECG signals, and accelerometric signals, respectively representative of the electrical signal of the heart and the movements of these individuals, associated with data on the start and end of seizures of these individuals, a step of partitioning (120) said learning base into a plurality of groups of individuals, called clusters (20a, 20b, 20c, 20d), each cluster being made up of individuals presenting the same type of epileptic symptom(s) in the vicinity of the epileptic seizures that they undergo, for each cluster (20a, 20b, 20c, 20d) made up of the previous step, a training step (130) of a machine learning model specific to this cluster,from meta-indicators derived from the data of said learning base restricted to the individuals belonging to this cluster, said meta-indicators being calculated by the following steps: o localization (131) of all the R peaks in the ECG signal of each individual of the cluster, an R peak being the most prominent peak of a QRS complex generated by a heartbeat of the individual, o identification (132) of the RR intervals of the ECG signal of each individual of the cluster, o calculation of indicators (133), called cardiac indicators, to, from the RR intervals including an average, a standard deviation, a low and high frequency power, the cardiac indicators being calculated at a predetermined time interval on the ECG signals of each individual, o calculation of indicators (134), called activity indicators, from the accelerometric signals of each individual of the cluster, the activity indicators being calculated at a predetermined time interval on the accelerometric signals of each individual, o operations (135) on said indicators calculated in the preceding steps to provide said meta-indicators representative of a distribution of the indicators within a predetermined time window, of each cardiac and activity indicator calculated, a step of classification (140) of said patient in a cluster chosen from said plurality of clusters so as to assign to said patient the corresponding trained learning model,a step of acquiring (150) ECG and accelerometric signals on said patient equipped with an ECG sensor and an accelerometer, a step of calculating said meta-indicators following the same principle as that implemented for training the learning models, a step of calculating (160) a crisis risk score for said patient, by feeding said trained learning model corresponding to the cluster of said patient with said meta-indicators calculated for this patient., 2. Method according to claim 1, characterized in that said step of classifying (140) said patient in a cluster comprises: a step of acquiring (141) ECG and accelerometric signals on said patient from an ECG sensor and an accelerometer worn by said patient, a step of calculating (142) said meta-indicators following the same principle as that implemented for training the models learning, a step of reporting (143) by said patient the first N attacks suffered, N being a predetermined integer, a step of calculating the risk score from each of the trained learning models, and of selecting (144) said trained model which provides the best risk score for all of the first N attacks reported, said patient being classified in the cluster corresponding to this model.

3. Method according to one of claims 1 or 2, characterized in that it further comprises a step of signifying (180) a crisis prediction error by said patient and a step of updating the learning model of said patient from said learning base restricted to individuals belonging to the cluster of said patient, to which said patient is added.

4. Method according to claim 3, characterized in that said step of updating the learning model of said patient comprises an allocation to the data from said patient of weights greater than the weights of the data of the other individuals of said learning base.

5. Method according to one of claims 1 to 4, characterized in that said meta-indicators representative of a distribution of the indicators within a predetermined time window include in particular an average, a variance, a minimum, a maximum of the indicators within this window.

6. Method according to one of claims 1 to 5, characterized in that said step of partitioning (120) said learning base into a plurality of clusters comprises: a step of randomly distributing said individuals of said learning base into K clusters, K being a predetermined integer, for each of the K clusters thus formed, a step of training a machine learning model specific to this cluster according to the same principle as that implemented in the context of training learning models from said meta-indicators derived from the data of the individuals of this cluster, a reassignment of the individuals from said learning base to the cluster whose learning model minimizes a prediction error, a reiteration of the previous steps until said total prediction error stabilizes around a value.

7. Method for assisting in the detection and / or prediction of epileptic seizures of an individual, called a patient, said method comprising: a method for determining an epileptic seizure risk score of said patient according to one of claims 1 to 6, a step of detecting and / or predicting (170) an epileptic seizure by comparing said seizure risk score determined by said determination method with at least one predetermined threshold.

8. System for determining an epileptic seizure risk score of an individual, called a patient, comprising: a learning base of epileptic individuals (20), said base comprising, for a plurality of individuals, electrocardiogram signals, called ECG signals, and accelerometric signals, respectively representative of the heartbeats and movements of these individuals, associated with data on the start and end of seizures of these individuals, a module for partitioning said learning base into a plurality of groups of individuals, called clusters (20a, 20b, 20c, 20d), each cluster being made up of individuals presenting the same type of epileptic symptom in the vicinity of the epileptic seizures that they undergo, for each cluster (20a, 20b, 20c, 20d) thus formed, a machine learning model specific to this cluster,trained from meta-indicators derived from the data of said learning base restricted to individuals belonging to this cluster, said meta-indicators being provided by a processing unit comprising:, o a module for locating all the R peaks in the ECG signal of each individual in the cluster, an R peak being the most prominent peak of a QRS complex generated by a heartbeat of the individual, o a module for identifying the RR intervals of the ECG signal of each individual in the cluster, o a module for calculating indicators, called cardiac indicators, from the RR intervals including a mean, a standard deviation, a low and high frequency power, the cardiac indicators being calculated at a predetermined time interval on the ECG signals of each individual, o a module for calculating indicators, called activity indicators, from the accelerometric signals of each individual in the cluster, the activity indicators being calculated at a predetermined time interval on the accelerometric signals of each individual,o a module for combining said calculated indicators to provide said meta-indicators representative of a distribution of the indicators within a predetermined time window, for each calculated cardiac and activity indicator, a module for classifying said patient in a cluster chosen from said plurality of clusters defined by the partitioning step, so as to select the corresponding trained learning model, a box (14) worn by said patient comprising at least one ECG sensor and an accelerometer, a module for calculating said meta-indicators according to the same principle as that implemented for training the learning models, a module for calculating a crisis risk score for said patient, by feeding said learning model corresponding to the cluster, of said patient with said meta-indicators calculated for this patient.

9. System for assisting in the detection and / or prediction of epileptic seizures of an individual, called a patient, comprising: a system for determining a seizure risk score according to claim 8, a module for detecting and / or predicting an epileptic seizure by comparing the risk score determined by said system for determining a seizure risk score with at least one predetermined threshold.

Images