SYSTEM AND METHOD FOR PREDICTING AND DETECTING EPILEPTIC SEIZURES

Abstract

SYSTEM AND METHOD FOR PREDICTING AND DETECTING EPILEPTIC SEIZURES The invention relates to a system for predicting and detecting an individual's epileptic seizures. Figure for the abstract: Figure 1

Description

Title of the invention: SYSTEM AND METHOD FOR PREDICTING AND DETECTING EPILEPTIC SEIZURES Technical field of the invention

[0001] The invention relates to a system and method for predicting and detecting epileptic seizures. Technological background

[0002] Epilepsy is a disabling neurological disease that affects tens of millions of people worldwide. Despite therapeutic progress, 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 significantly higher than that of the general population is observed.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.

[0003] Epilepsy is a complex and plural disease characterized by variable etiologies, ages of onset and clinical manifestations. The International League Against Epilepsy proposes a classification of seizures based in particular on the location of the initial manifestations (focal or generalized), the level of consciousness, the motor signs or even the possible propagation of the seizure. All seizures can be prolonged and lead to status epilepticus, which represents a major medical emergency. Epilepsy can also be accompanied by cognitive or psychiatric comorbidities.

[0004] In the particular 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 the seizures constitutes a source of considerable anxiety. The quality of life of these people and their loved ones is consequently significantly degraded.

[0005] The measurement of seizure frequency is an indicator considered essential for estimating the quality of care. Several studies have highlighted the difficulty that maintaining a reliable seizure diary represents for patients. phenomena such as memory loss or altered perception at the time of the seizure are thought to be involved. 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. Such systems would limit the risk of seizure-related accidents, provide some reassurance to patients and their loved ones on a daily basis, and ultimately improve the quality of care provided by the neurologist.

[0006] There are already commercially available systems for detecting certain types of seizures which notably present significant motor manifestations such as generalized tonic-clonic seizures. These are particularly worrying seizures because they can lead to sudden unexplained deaths.

[0007] Several approaches have been tested and they are generally based on the monitoring of physiological signs associated with the search for biomarkers and coupled with detection algorithms. The signs monitored are as diverse as EEG (various montages), video, audio, movements, heart rate, skin conductance, skin temperature, oxygenation rate, respiration or even muscular activity via electromyogram. Multimodal approaches, i.e. associating several sensors at the same time, are considered the most promising.

[0008] However, there is currently no non-invasive system that can reliably detect focal seizures.

[0009] Beyond detection, anticipation of crises could truly change the lives of drug-resistant patients.

[0010] The inventors therefore sought to develop a new system and method which makes it possible to detect and anticipate crises in order to improve the monitoring and comfort of patients and their loved ones. Objectives of the invention

[0011] The invention aims to provide a system and a method for predicting and detecting non-invasive epileptic seizures. Statement of the invention

[0012] To do this, the invention relates to a system for predicting and detecting epileptic seizures in an individual comprising the following technical characteristics: - a box configured to be able to be worn by said individual, said box comprising at least one sensor, called ECG, configured to acquire data representative of the electrocardiogram of the individual and at at least one accelerometer configured to provide acceleration measurements of the case worn by the individual, - a unit for processing the measurements provided by the various sensors of said box, said processing unit comprising at least the following modules: • a module for calculating data, called cardiac data, from the electrocardiogram data, representative of the temporal and frequency variations of the electrocardiogram data, • a data calculation module, called activity data, from the measurements provided by said accelerometer and representative of the energy expenditure, posture and / or movements of the individual, • a module for calculating data, called a crisis risk score, based in particular on said cardiac data and said activity data, said module for calculating the risk score implementing a trained machine learning model to determine a crisis risk score, the model having been trained using a training database which includes activity data and cardiac data from a plurality of individuals associated with occurrences of detected crises, • a crisis reporting module by comparing the risk score to a predetermined threshold, depending on the individual.

[0013] The system according to the invention therefore 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. For this, the system according to the invention assigns a group to the patient based on a comparison of epileptic data provided by the patient with pre-recorded data. The system according to the invention thus implements a machine learning model per group of individuals.

[0014] The applicant used the European epilepsiae database to implement and test the system according to the invention.

[0015] This database is made up of recordings made in 3 European centers (France - La Pitié Salpêtrière, Germany - Freiburg, Portugal - Coimbra). The data of 168 patients (out of the 275 patients that make up the database

[0016] complete) were recovered by the applicant to develop the system and method according to the invention. This database includes several days of data per patient (8 on average, 24 maximum, 3 minimum).

[0017]

[0018]

[0019]

[0020]

[0021]

[0022]

[0023]

[0024]

[0025]

[0026] The following data were used for each selected patient: - ECG data, - age, date of admission, neurological assessment, treatment, assessments, frequency of seizures, main type of seizures - Metadata for each seizure: start and end of the seizure on the EEG, start and end of the seizure from a clinical point of view (video), type of seizure, - seizures on average per patient, 94 maximum, 2 minimum. The applicant also equipped patients with a belt known as BioHamess 3.0 from the Zéphyr® brand (Medtronic®). This equipment allows continuous collection of an electrocardiogram (ECG) (sampled at 1 kHz), 3D accelerometry data (sampled at 100 Hz), and rib cage deformation data to report respiration (sampled at 25 Hz). The depositor carried out data measurements for an average of 17 days (maximum 21 days, minimum 12). During this period, an average of 23 seizures were detected per patient, a maximum of 107 and a minimum of 2. 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. These changes are not the same in all patients (for example, some have tachycardia, others bradycardia). Similarly, in some patients, seizures are accompanied by gestures (or the absence of gestures) which can be captured by acceleration sensors if available. Given this diversity of manifestations, it was determined that it was necessary to have a model specific to each type of manifestation to improve the quality of detection. Thus, rather than having a model specific to all patients, the invention provides for assigning the patient to a group of individuals from among a plurality of groups and having a machine learning model per group thus constituted. It was therefore planned to learn a model for each group of patients exhibiting the same types of symptoms around the time of attacks. The automatic learning of the model for each group is done prior to its use within the system. Throughout the following, this pre-learning phase is determined offline. For a new patient who is equipped with the system according to the invention, the first step is to determine which group he or she belongs to. This phase is referred to throughout the following as the online pre-learning phase (or group assignment phase).

[0027] This step aims, from a reduced data set, to determine the patient's group membership. These data are data from previous attacks reported by the patient himself, which makes it possible to select the typology of symptom that he presents and thus the most appropriate model to use.

[0028] The model thus selected makes it possible to initialize a patient-specific model. The system can operate in continuous mode. The patient model is not fixed: as soon as the patient gives feedback on a detection error or a missed seizure, the system can change the group to which it belongs.

[0029] The system according to the invention comprises a box worn by the patient, this box comprising several sensors including an ECG sensor and an accelerometer. These sensors communicate, for example via Bluetooth, with an application embedded on a smartphone.

[0030] This application compresses data and sends it to a remote server, at a configurable time interval.

[0031] The application incorporates the most up-to-date patient learning model (its parameters are therefore fixed in the application): this runs continuously for the real-time alert system.

[0032] Regularly (about 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 patients provide.

[0033] The principle of the learning phase (off-line) of the model applicable to the patient is as follows: Preparing ECG data

[0034] All R peaks in the ECG signal are localized. This is the most prominent peak in the characteristic wave called the QRS complex that a heartbeat generates. The distance between two consecutive R peaks is conventionally used to determine the instantaneous heart rate.

[0035] The applicant has developed his own algorithm for this so that the detection is robust to noise. This algorithm also makes it possible to automatically identify portions of signals that are too noisy for the R peaks to be detectable. This algorithm implements a hidden Markov model.

[0036] The series of intervals between two R peaks is called the RRI series or tachogram. Feature extraction

[0037] The system calculates a number of indicators from the RR series, including the following indicators: - mean, standard deviation, norm ^^2 of the derivative (described for example in the

[0038]

[0039]

[0040]

[0041]

[0042]

[0043]

[0044]

[0045]

[0046]

[0047] document Malik et al. 1996) - low and high frequency power (described for example in Malik et al. 1996) - Cardiac sympathetic index (CSI), modified CSI (described for example in the document Jeppesen et al. 2019) - Sample entropy, Detrended Fluctuation Analysis, Deceleration & Acceleration capacity, Horizontal Visibility Graph (described for example in the document Thu et al. 2019). The system also uses indicators from acceleration signals: - posture, - energy expenditure, - physical activity All these indicators are calculated over sliding windows whose size can vary, thus multiplying their number. For example, the system calculates an indicator value every 10 seconds. Finally, every 10 seconds, the system retrieves all indicator values ​​from the last 10 minutes and calculates meta-indicators for each indicator (mean, variance, minimum, maximum, etc.) It is these meta-indicators that feed the supervised learning model. Added to this data is an a priori crisis probability indicator. This is calculated based on the history of crises and any cycles observed in patients. Alternatively, it is possible to add metadata corresponding to the patient (epileptic syndrome, etc.). Learning model Patients were grouped by symptom type. Indicators of patients with similar symptoms will result in a specific model. This can be done automatically by finding the clusters that minimize the total classification error. Alternatively, it is possible to draw inspiration from the principle of the k-means algorithm, the principles of which are as follows: - we fix a number K of clusters and a random initialization is done, - for each cluster, a model is trained, - once the cluster models are trained, patients are reassigned by associating them with the cluster whose model minimizes the prediction error; - the process is iterated several times until the total prediction error stabilizes. Alternatively, SVMs (Support Vector Machines) or Random Forests are used for binary classification models, applied to the previously mentioned indicators.

[0048] It is also possible to use deep neural network models, applied either to these same indicators or directly to the RRI series + a priori probability indicator.

[0049] It is also possible, according to one embodiment, to complicate the models in order to classify the type of epileptic seizures (focal seizure, generalized seizure, etc.).

[0050] The principle of use of the device is as follows. - When a patient equips himself with the device according to the invention, a rapid pre-learning phase is implemented to assign him a specific group, - The data provided by the sensor are collected in real time, and the same indicators as those used in the offline pre-learning phase are calculated by the modules of the processing unit, - The patient can also be asked to report the first N seizures via an application or their personal space (web interface), where N is a predetermined integer. This allows a pre-learned model that best suits the patient to initialize the patient-specific model and to predict future seizures. - At any time the patient can report a prediction error. If this is the case, his model is updated. - In order to avoid over-learning, updating a patient model is done by including it in the patient cluster of the pre-learning database which was used to pre-learn the model. Detailed description

[0051] The system and method according to the invention was tested by monitoring in real life three physiological parameters (ECG, respiration and 3-axis accelerometry) in pharmacoresistant epileptic patients using simple instrumentation (sensors) and to verify that the seizure detection by the method according to the invention is consistent with the seizure detection by practitioners monitoring these patients.

[0052] Thus, the data collected using the sensors are compared with the data collected on a crisis diary (standardized format already used in current practice) completed in real time by the healthcare staff who continuously support the patients. These crisis diaries are completed during the periods when the patients wear the sensor.

[0053] As indicated, according to one embodiment, BioHarness 3 Zephyr sensors are used.

[0054] The heart rate (HR sensor), respiration (Breathing Sensor) and actimetry sensors are all located on the belt. The device is switched on and off by pressing the box located under the left armpit.

[0055] The belt does not emit an audible alarm but indicator lights located on the box under the left armpit indicate whether the cardiac signal is being correctly captured. A continuously lit red light indicates that the cardiac signal is not being captured correctly, a flashing red light indicates that the cardiac signal is being correctly collected.

[0056] A different colored indicator light indicates whether the battery needs to be recharged. To recharge the battery, the box under the armpit is removed and placed on a charging base connected to a computer. A recharge lasts between 1 and 3 hours. It is therefore possible for each patient to have two boxes that the caregivers will alternate continuously to avoid losing the signal. A charged box has an autonomy of approximately 30 hours.

[0057] The analysis consists of qualifying the separability of the signals according to two classes: patient in normal state (interictal) and in crisis state (ictal or peri-ictal). A preliminary step of identifying the noise in the signals (due to movements, muscular artifacts linked or not to a crisis, to the bad positioning of the sensor, etc.) then of cleaning or possibly rejecting the data which is too noisy is carried out.

[0058] The QRS complexes (or R peaks) are identified from the ECG tracings, from which the following will be deduced:

[0059] - heart rate,

[0060] - heart rate variability (HRV signal).

[0061] Temporal, frequency and geometric / non-linear variables are calculated from the HRV signal. The HRV signal can optionally be corrected upstream (detection and correction of ectopic beats, detection and correction of non-physiological / aberrant heart rate).

[0062] The respiratory rate as well as the amplitude of inspirations / expirations are calculated from the raw respiration trace.

[0063] An estimate of energy expenditure is calculated from the accelerometer traces.

[0064] The statistical properties of these different variables are measured and described according to the patient context: patient in normal state (inter-ictal), or in crisis state (ictal). Different measures of distances and indices (example: Dunn Index) are used to qualify the separability between the interictal and ictal (or peri-ictal) traces.

[0065] The occurrence of epileptic seizures is evaluated using the collection of two complementary data: physiological data of interest from the sensors directly (ECG, respiration and accelerometers), and seizure characteristic data. collected in real time using a seizure diary kept by the nursing staff (who are permanently present with the patients). The different seizure detection thresholds from the sensor data are determined by maximizing the percentage of positive agreement (= Recall) and the precision (defined as the ratio of the number of true positives to the sum of the number of true positives and false positives, which corresponds to the positive predictive value).

[0066] Thus, four situations can arise:

[0067] - True positives which are the actual crises both detected by the system according to the invention (variation of physiological data of interest around the time of the crisis) and confirmed by the recording of the crisis on the diary by the nursing staff,

[0068] - False negatives are the actual seizures observed and recorded on the diary of crises but not detected by the system according to the invention,

[0069] - False positives also called false alerts are situations where the sensors record variations but where no observable crisis is collected in the crisis diary.

[0070] - True negatives which correspond to the normal (healthy) state of the patient where the sensors do not detect any variation and no observable crises are collected in the crisis diary. This data is not quantifiable.

[0071] Crisis detection methods are interpreted according to two main criteria

[0072] - the percentage of positive agreement, defined as the ratio of the number of true positives on the sum of the number of true positives and false negatives,

[0073] - the number of false alert(s) per hour (or per day), this criterion being evaluated on interictal periods exclusively.

[0074] Three physiological variables are measured by the BioHamess 3 Zephyr sensor:

[0075] - the electrocardiogram (ECG) - 1000 Hz

[0076] - breathing - 100 Hz

[0077] - 3-axis accelerometer (X, Y, Z) - 100 Hz / axis.

[0078] Regarding retrospective EEG-video examinations, ECG tracings (sampling frequency may vary depending on the measuring tool) as well as oxygen saturation (SpO2) will be collected.

[0079] Seizure diaries (prospective tracings) and seizure start and end annotations (tracings from retrospective EEG-video examinations) kept by caregivers will be collected.

[0080] EEG-video tracings: This examination is the golden rule for the diagnosis and monitoring of epilepsy. Each patient included has therefore already undergone at least one EEG-video examination in the past.

[0081] Analysis methodology: within the investigation center, the EEG-video traces are annotated by the neurologist (identification of the start and end time of the seizure based on EEG signals and video) using EEG trace replay software. This annotation data is then integrated into the system's learning base.

Claims

1. Claims System for predicting and detecting epileptic seizures in an individual, called a patient, said system comprising: • a module for assigning to said patient a group chosen from a plurality of predetermined patient groups, each representative of a type of symptoms appearing in the vicinity of the attacks, • a box configured to be worn by said patient, said box comprising at least one ECG sensor configured to acquire data representative of the individual's electrocardiogram and at least one accelerometer configured to provide acceleration measurements of the box worn by the individual, • a unit for processing the measurements provided by the various sensors of said box, said processing unit comprising at least the following modules: • a data calculation module, called cardiac data, from electrocardiogram data, representative of the temporal and frequency variations of the electrocardiogram data, • a data calculation module, called activity data, from the measurements provided by said accelerometer and representative of the energy expenditure, posture and / or movements of the individual, • a module for calculating data, called a crisis risk score, as a function of said cardiac data and said activity data, said risk score calculation module implementing a trained machine learning model to determine a crisis risk score, the model having been trained using a training database which includes activity data and cardiac data of a plurality of individuals belonging to said group of said patient, associated with occurrences of detected crises, • a crisis reporting module by comparing the risk score to a single predetermined, function of said individual.

2. System according to claim 1, characterized in that said module for assigning a group to said patient comprises means for receiving seizure data provided by said patient, for comparing these received data with pre-recorded data representative of the different groups and for selecting the most representative group.

3. Method for predicting and detecting epileptic seizures of an individual, called a patient, said method comprising: • a preliminary step of assigning to said patient a group chosen from a plurality of predetermined groups of individuals, each representative of a type of symptoms appearing in the vicinity of the seizures, • a step of acquiring measurements provided by an electrocardiogram worn by said patient • a step of acquiring measurements provided by an accelerometer worn by said patient, • the processing of said measurements provided by said electrocardiogram and by said accelerometer, this processing comprising: • the calculation of data, called cardiac data, from the electrocardiogram data, representative of the temporal and frequency variations of the electrocardiogram data, • the calculation of data, called activity data,from the measurements provided by said accelerometer and representative of the energy expenditure, posture and / or movements of the individual, • the calculation of data, called a crisis risk score, based on said cardiac data and said activity data, said risk score calculation module implementing a machine learning model trained to determine a crisis risk score, the model having been trained using a training database which includes, activity data and cardiac data of a plurality of individuals, associated with detected seizure occurrences, said plurality of individuals belonging to the group of said patient, a step of reporting a crisis by comparing the risk score to a single predetermined, function of said individual.

Images