Method for predicting the risk of sudden death and related device

Abstract

The present invention relates to a method for predicting the risk of occurrence of sudden death in a patient, the method being computer implemented and comprising the steps of: - receiving data on patient care pathways; - determining from the received data whether the patient belongs to one of a set of predefined groups, each group being associated with a set of predefined data; - for each determined predefined datum, searching for predefined datum values ​​relating to that patient to obtain a set of values ​​specific to that patient; applying the neural network to patient-specific values, specific to the group for which the neural network has been determined, to obtain the patient's risk of occurrence of sudden death.

Description

[Technical field]

[0001] This patent application claims the benefit of document FR21 / 12396, filed November 23, 2021, which is incorporated herein by reference.

[0002] TECHNICAL FIELD OF THEINVENTION The present invention relates to a method for predicting the risk of occurrence of sudden death in a patient. The present invention also relates to a computer program product and a readable information carrier for implementing the method. [Background technology]

[0003] Technical Background of the Invention Sudden death is defined as an unexpected death without obvious non-cardiac causes, occurring within one hour after the onset of symptoms, with or without bystanders, and occurring rapidly from debilitating events. This condition affects 30,000 to 40,000 people per year in France and 300,000 people per year in Europe.

[0004] The prognosis remains extremely poor, with survival rates below 10% in some recent studies. Several tools have been proposed to improve prognosis, including early bystander cardiac massage, early defibrillation, or hospital management (via early coronary artery management or application of therapeutic hypothermia), especially in terms of prehospital management, via the chain of survival.

[0005] Nevertheless, survival outcomes remain disappointing, although some studies have shown recent improvements.

[0006] Given these modest therapeutic outcomes, several prophylactic alternatives have been proposed to prevent the occurrence of such events. Thus, the development of antiarrhythmic therapy and automatic implantable cardioverter-defibrillators has allowed significant prevention in patients determined to be at high risk for sudden death.

[0007] Therefore, optimizing the use of these preventative therapies depends on identifying at-risk patients.

[0008] Although much research has been done on the "post-mortem" aspects in this context, especially on the care of patients who die suddenly, predicting the occurrence of such events remains difficult, so identifying patients at risk remains a major research challenge, with disappointing results so far.

[0009] Certain patient groups at very high risk of sudden death (certain structural or electrical arrhythmogenic heart diseases) have been identified and are already receiving specialized rhythmology management.

[0010] However, from an epidemiological point of view, such a component represents a very small portion of the total population, and the majority of patients who die suddenly are not part of such a population. In fact, the main cause of sudden death remains ischemic heart disease, which occurs either during the acute event (myocardial infarction) or during follow-up in these patients. The cohort of patients with ischemic heart disease is very large, and only a very small proportion of them experience sudden death during the course of the disorder.

[0011] Thus, there is a mismatch between the small but very high-risk groups (certain arrhythmogenic, structural, and electrical heart diseases) and the large but low-risk groups (ischemic heart disease), who therefore constitute the majority of sudden deaths in the general population.

[0012] Thus, predicting sudden death remains a challenge, as most patients cannot benefit from individual risk stratification because population risk factors are not clearly identified (unlike, for example, global cardiovascular risk, where individual risk assessment is possible with tools such as the Framingham score).

[0013] To stratify this individual risk, several risk factors, including family history, have been proposed. In the population of patients with ischemic heart disease, which represents the majority of sudden death victims, prediction is currently based primarily on left ventricular ejection fraction, with relatively disappointing results. Summary of the Invention

[0014] Therefore, there is a need for a method to predict the risk of occurrence of sudden death in a patient.

[0015] To this end, the present specification describes a method for predicting the risk of occurrence of sudden death in a patient, the method being implemented by a computer and comprising the steps of receiving data relating to a patient's care pathway, determining from the data relating to the care pathway whether the patient belongs to one of a set of predefined groups, each group being associated with a set of predefined data, searching for values ​​of the predefined data for the patient for each of the determined predefined data to obtain a set of patient-specific values ​​to obtain the determined groups and the determined set of predefined data, and applying a neural network to the patient-specific values, the neural network being specific to the determined groups, to obtain the risk of occurrence of sudden death in the patient.

[0016] This method is distinct from machine learning methods for analyzing electrocardiogram signals or populations at high cardiovascular risk, and indeed such methods are limited to populations at cardiovascular risk.

[0017] In contrast, the present invention proposes predictions not only for those at cardiovascular risk but also for the general population: since only 5% of the population is at cardiovascular risk, it is necessary to be able to predict the risk of sudden death for the remaining 95%.

[0018] This led to the identification of new groups, the eight groups referred to in this application, none of which have been so identified to date.

[0019] For example, the present invention makes it possible to know whether someone is at increased risk of sudden death if they do not have their teeth cleaned annually or if they are taking psychotropic drugs.

[0020] This method allows us to identify people who are likely to die suddenly and who need implantable cardioverter-defibrillators. These people are not generally considered to be at risk for sudden death because they are not part of the cardiovascular risk population. However, they represent 90% of the total population at risk for sudden death.

[0021] The present process therefore makes it possible to effectively combat the scourge of sudden death.

[0022] According to particular embodiments, the prediction method exhibits one or more of the following characteristics, taken alone or in any technically possible combination: - Each neural network is a gated recurrent neural network. - The data received is about the patient's care pathway for the last 5 years. Each given datum is the presence of a disorder or the intake of a medication. the number of predetermined data of the group is between 10 and 30, preferably between 15 and 25 and advantageously equal to 20; - the group is associated with certain data on respiratory diseases or the consumption of products limiting respiratory diseases; - the group is associated with certain data concerning a neurological disorder or the intake of products limiting a neurological disorder; - the group is associated with certain data concerning cancer disease or the intake of products limiting cancer disease; - the group is associated with certain data concerning addictive disorders or the consumption of products limiting addictive disorders; - the group is associated with certain data concerning ageing disorders or the consumption of products limiting ageing disorders; - the group is associated with certain data concerning heart disease or the consumption of products limiting heart disease; - the groups are obtained by applying a k-means partitioning technique to a set of data comprising data on the care pathways of a set of patients, and each given data of the group is obtained by applying a linguistic analysis tool to the set of data comprising data on the care pathways of the patient. - the set of data comprises artificial data generated by applying a function to data relating to care pathways of a set of patients. - The function is an adversarial neural network.

[0023] The present specification also describes a computer program product comprising program instructions forming a computer program stored on a readable information medium, the computer program being loadable into a data processing apparatus and which, when the computer program is executed on the data processing apparatus, performs the evaluation method as described above.

[0024] The present specification also relates to an information readable medium comprising program instructions forming a computer program, which is loadable into a data processing apparatus and which, when the computer program is executed on the data processing apparatus, performs the evaluation method as described above. [Brief description of the drawings]

[0025] The characteristics and advantages of the invention will become apparent from the following description, given by way of non-limiting example only and made with reference to the accompanying drawings, in which: [Figure 1]1 is a schematic diagram of a system and a computer program product. [Diagram 2] 1 is a flow chart of an embodiment of a method for predicting the risk of occurrence of sudden death in a patient. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0026] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Description of the system used A system 10 and a computer program product 12 are shown in FIG.

[0027] The interaction between the system 10 and the computer program product 12 enables the implementation of a method for predicting the risk of occurrence of sudden death in a patient. The prediction method is therefore a computer-implemented method.

[0028] System 10 may be a desktop computer. Alternatively, system 10 may be a rack-mounted computer, a laptop computer, a tablet, a personal digital assistant (PDA), or a smartphone.

[0029] In FIG. 1, system 10 includes a computer 14 , a user interface 16 , and a communication device 18 .

[0030] Computer 14 is electronic circuitry designed to manipulate and / or convert data represented by electronic or physical quantities in registers and / or memories of system 10 into registers or other similar data that correspond to the physical data in the memory of another type of display, transmission, or storage device.

[0031] As specific examples, computer 14 may include single-core or multi-core processors (such as central processing units (CPUs), graphics processing units (GPUs), microcontrollers, and digital signal processors (DSPs)), programmable logic circuits (such as application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), programmable logic devices (PLDs), programmable logic arrays (PLAs)), state machines, logic gates, and discrete hardware components.

[0032] The computer 14 includes a data processing device 20 capable of processing data, particularly by performing calculations, a memory 22 capable of storing data, and a reader 24 capable of reading a computer-readable medium.

[0033] The user interface 16 includes input devices 26 and output devices 28 .

[0034] Input device 26 is a device through which a user of system 10 inputs information or commands into system 10 .

[0035] 1, the input device 26 is a keyboard. Alternatively, the input device 26 is a pointing device (mouse, touchpad, graphic tablet, etc.), a voice recognition device, an eye tracker, or a haptic (motion analysis) device.

[0036] Output device 28 is a graphical user interface, i.e., a display device designed to present information to a user of system 10.

[0037] 1, output device 28 is a display screen for visually presenting output, in other embodiments output device 28 is a printer, an augmented and / or virtual display device, a speaker or other audio generating device for presenting output in audio form, a vibration and / or odor generating device, or a device capable of generating an electrical signal.

[0038] In one particular embodiment, the input device 26 and the output device 28 are the same component forming a human-machine interface, such as an interactive screen.

[0039] The communication device 18 allows for one-way or two-way communication between the components of the system 10. For example, the communication device 18 is a bus communication system or an input / output interface.

[0040] The presence of communications device 18 allows components of computer 14 to be remote from one another in some embodiments.

[0041] The computer program product 12 includes a computer-readable medium 30 .

[0042] Computer-readable medium 30 is any tangible device that can be read by reader 14 of computer 14 .

[0043] In particular, computer readable medium 30 is not itself a ephemeral signal such as an electric wave or other freely propagating electromagnetic wave such as a light pulse or an electronic signal.

[0044] Such computer-readable storage medium 30 may be, for example, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any combination thereof.

[0045] As a non-limiting list of more specific examples, computer readable storage medium 30 may be a mechanically encoded device such as a punch card or groove embossed structure, a floppy disk, a hard disk, a read only memory (ROM), a random access memory (RAM), an erasable programmable read only memory (EPROM), an electrically erasable programmable read only memory (EEPROM), a magneto-optical disk, a static random access memory (SRAM), a compact disk (CD-ROM), a digital versatile disk (DVD), a USB flash drive, a floppy disk, a flash memory, a solid state drive (SSD), or a PC card, such as a PCMCIA memory card.

[0046] The computer program is stored on a computer readable storage medium 30. The computer program comprises one or more sequences of stored program instructions.

[0047] Such program instructions, when executed by the data processing apparatus 20, effect the steps of the estimation method to be performed.

[0048] For example, the program instructions may be in source code form, computer executable form, or any intermediate form between source code form and computer executable form, such as a form resulting from translation of source code via an interpreter, assembler, compiler, linker, or localizer, etc. Alternatively, the program instructions may be microcode, firmware instructions, state definition data, integrated circuit configuration data (e.g., VHDL), or object code.

[0049] The program instructions may be written in any combination of one or more languages, such as, for example, object-oriented programming languages ​​(FORTRAN, C++, JAVA, HTML), procedural programming languages ​​(eg, C), and the like.

[0050] Alternatively, the program instructions may be downloaded, as in the case of an application, from an external source over a network, in which case the computer program product comprises a computer readable data carrier on which the program instructions are stored, or a data carrier signal in which the program instructions are encoded.

[0051] In any case, the computer program product 12 comprises instructions which can be loaded into the data processing apparatus 20 and which, when executed by the data processing apparatus 20, cause the execution of the foreseen method. According to the present embodiment, the execution is carried out wholly or partly in the system 10, i.e. either in a single computer, or in a distributed system among several computers (in particular through the use of cloud computing).

[0052] Operation of the system 10 will now be described with reference to FIG. 2, a flow chart illustrating an embodiment of a method for prediction.

[0053] The method is a method for predicting the risk of occurrence of sudden death in a patient.

[0054] As a non-limiting example, the patient is an adult human patient.

[0055] The human subject may have any medical profile. The subject may not be at particular cardiovascular risk.

[0056] The method is intended to be applicable to any type of population.

[0057] The method comprises two stages: a preparation stage P1 and an operation stage P2.

[0058] Depending on the case, the preparation step P1 may be performed by the system 10 or by another system. Typically, the preparation step P1 is performed well before the implementation step P2.

[0059] In this case, as will be explained below, the preparatory stage P1 comprises three steps: a training step E50, a decision step E52 and a training step E54.

[0060] We now describe the operational phase P2, which includes a receiving step E56, a determining step E58, a searching step E60 and an applying step E62.

[0061] In a receiving step E56, the system 10 receives data relating to the patient's care history over the past five years.

[0062] Thus, patient data is care data.

[0063] Data on the care pathway describes the length of the patient's stay, the disorders found in the patient, the tests performed and the treatments applied.

[0064] In France, this data is available through public authorities, usually social security or local health agencies.

[0065] Nevertheless, it may also be envisaged that data regarding care pathways is obtained using answers from the patient to previous care pathway questions.

[0066] The five-year period was chosen because applicants have demonstrated that shorter periods result in less reliable projections, and that longer periods do not result in more reliable projections.

[0067] Here, it can be specified that the data relating to the health course is not recorded signal data such as electrocardiogram signals, etc. Only the results of the interpretation are taken into account.

[0068] Moreover, health journey data is broader than the collection of recorded heart disease signals.

[0069] In particular, data about health journeys includes data about any type of pathology. For example, health journey data can indicate when scaling occurred, as well as test results.

[0070] Furthermore, data on health pathways are heterogeneous in the sense that they bring together data of different natures.

[0071] Thus, data regarding a health course may include data selected from classes such as medications taken, medical history, doctor visits, pathology list, laboratory results, firefighter interventions, emergency room visits, etc.

[0072] Preferably, the data regarding the health course includes all previous classes.

[0073] During a decision step E58, the system 10 searches among a set of predetermined groups for the group to which the patient belongs.

[0074] To do this, the system 10 applies a classification function to the data received in the receiving step E50 to determine the closest group.

[0075] These groups are predefined groups that exhibit relatively similar behavior with respect to risk of sudden death occurrence.

[0076] The groups were obtained by applying a k-means partitioning technique to a set of data containing care pathway data for a set of patients.

[0077] The data set is representative of the general population in Applicants' experiments.

[0078] The data set also contained sufficient data to allow for grouping on the occurrence of sudden death.

[0079] In the presence of insufficient or unbalanced samples (low number of sudden death cases), the data set includes artificial data generated by applying a function to data on the care pathways of a set of patients.

[0080] According to a particular example, this function is an adversarial neural network that can be generated from care pathway data.

[0081] Furthermore, each group is associated with a predefined set of data.

[0082] In the example described, each given datum is the presence of a disorder or the ingestion (administration) of a product.

[0083] Each given piece of data for a group is obtained by applying a linguistic analysis tool to a set of data that includes data relating to the care pathways of a set of patients.

[0084] The number of predetermined data of a group is between 10 and 30, preferably between 15 and 25, advantageously equal to 20.

[0085] More preferably, the amount of data of the type "presence of disorder" and the amount of data of the type "administered product" are equivalent.

[0086] According to the described embodiment, the predetermined group includes: - a group on prescribed data on respiratory diseases or on the intake of medicines limiting respiratory diseases, - a group on prescribed data on neurological disorders or on the use of drugs to limit neurological disorders, - a group on prescribed data relating to cancer diseases or the use of drugs to limit cancer diseases, - a group on prescribed data on addictive disorders or on the use of medications limiting addictive disorders, and - a group on prescribed data on ageing disorders or on the use of drugs to limit ageing disorders, - a group on certain data relating to cardiac diseases or the use of drugs to limit cardiac diseases, Includes.

[0087] In some cases, it is conceivable that there may be multiple groups treating the same disease, and in particular, it is considered preferable to have two groups treating heart diseases, as shown by the applicant's experience.

[0088] In either case, the groups and the predetermined data are obtained by carrying out a receiving step E50 of forming a database, and a step E52 of determining the groups and the predetermined data.

[0089] During a search step E60, the system 10 searches, for each determined predetermined datum, the values ​​of the predetermined datum for the patient, to obtain a set of patient-specific values.

[0090] Such searches can be reduced, for example, for extraction of care pathway data.

[0091] The term "value" herein is understood broadly to include both a binary value (medication taken or not) or a quantitative value (usually a scale between 1 and 10 for pain).

[0092] Alternatively, or additionally, these values ​​may be obtained via questioning of the patient or caregiver.

[0093] In an application step E62, the system 10 applies the neural network, which is specific to the group for which the neural network has been determined, to the patient-specific values ​​in order to obtain the risk of occurrence of sudden death in the patient.

[0094] The neural network is notably pre-trained by using the same data that made it possible to find the given group, this corresponds to the training step E54 of the first stage P1.

[0095] This means that the system 10 has in memory the attributes of each given group, as well as the particular neural network.

[0096] These neural networks are unique in that their inputs are the values ​​of given data relating to the group under consideration.

[0097] Typically, for a group of given data relating to a respiratory disorder or the intake of a medication to limit the respiratory disorder, the neural network takes as input 10 values ​​of the given data relating to the respiratory disorder and 10 values ​​of the intake of a medication to limit the respiratory disorder.

[0098] Moreover, in the illustrated example, each neural network is a recurrent gated neural network.

[0099] The probability value is expressed as a score for the next three months, for example.

[0100] However, any form of score can be considered to give the calculation result of the neural network.

[0101] Therefore, this method can effectively predict the risk of occurrence of sudden death.

[0102] This will be introduced in the next section.

[0103] Experimental Results This method has been the subject of experimentation by applicants and is described below.

[0104] Purpose of the experiment The main objective of these experiments is to predict the occurrence of sudden death in adults by comparing data from case-patients (victims of sudden death) and four control populations. - Group 1: Patients with ischemic heart disease without sudden death, - Group 2: Patients with acute coronary syndrome without sudden death - Group 3: Patients with chronic heart failure without sudden death, and - Group 4: Individuals who are representative of the general population.

[0105] These studies used the register of the Centre d'Expertise de la Mort Subite (Sudden Death Expertise Center) and the medical administrative database of the Assurance Maladie (Health Care System).Their scientific interest was evaluated by the French Society of Cardiology, which highlights their public health interest and the great contribution they make.

[0106] These experiments sought to construct groups of patients within the sudden death population, identify and describe the heterogeneity of this population and their associated risk factors, and develop a prediction algorithm for sudden death based on the observed trajectory of care prior to the event.

[0107] Data used Study cohort description Included patients

[0108] Since May 2011, the Sudden Death Specialist Centre has been collecting all cases of sudden death occurring in a given geographical region (Paris and the three adjacent regions, Hauts-de-Seine, Seine-Saint-Denis and Val-de-Marne), representing a total population of 6.6 million inhabitants, i.e. 10% of the French population. This collection is made possible by a multilevel collaboration between pre-hospital emergency services (Paris Fire Brigade, SAMU), hospitals (Resuscitation and Cardiology Departments) and the Paris Institute of Legal Medicine.

[0109] For all included cases, information on occurrence of events (Utstein criteria), management (pre- and during hospitalization), and patient outcome (survival and neurological prognosis) was collected prospectively using multiple sources and frequent quality control (assessed to be comprehensive for up to 99% of cases in the area of ​​interest). This collection received a favorable opinion from the Comite consultatif sur le traitement de l'information en matiere de research (CCTIRS (Advisory Committee on the Treatment of Information in Research and Studies) file N12.336) and was approved by the CNIL (National Committee for Information and Liberty) (decision DR-2012-445).

[0110] Therefore, the case population of this study was CEMS registered patients who presented with sudden death between May 15, 2011 and December 31, 2020. Over the considered study period of more than 9 years (2011–2020), 24,000 cases were included.

[0111] Control group Control populations were defined and collected from the Assurance Maladie medical administrative database. These populations included four different control cohorts, with each case matched to three controls matched for sex, age, and place of residence. Thus, 288,000 controls were included in these studies.

[0112] For the first cohort (patients with ischemic heart disease), 72,000 controls were selected from this cohort to achieve a 3:1 matching of controls and cases. Identification of patients with ischemic heart disease was performed according to the medical mapping method described previously.

[0113] The same methodology was followed for the second (patients with acute coronary syndrome) and third (patients with chronic heart failure) populations.

[0114] For the fourth cohort (individuals representative of the general population), a control group representative of the general population was constructed, with three controls selected per case. A total of 72,000 controls were randomly selected from Paris and three adjacent regions (Hauts-de-Seine, Seine-Saint-Denis, and Val-de-Marne), excluding individuals previously included in the three predefined cohorts.

[0115] Description of the data used Within the framework of these studies, the medical histories of the 24,000 sudden death cases and 288,000 controls mentioned above were collected over a period between 5 and 10 years before the sudden death. This information was extracted from the medical administrative database of the Assurance Maladie of the SNDS (Systeme National des Données de Santé).

[0116] The SNDS is a repository of pseudo-anonymized healthcare administrative data covering the entire French population and including all care presented for reimbursement. It is managed by the Caisse Nationale de l'Assurance Maladie (CNAM) and links data from Assurance Maladie (SNIIRAM database), hospital data (PMSI database) and medical cause of death data (INSERM CepiDC database).

[0117] It currently contains over 3,000 variables, representing the flow of 1.2 billion medical events, 11 million hospitalizations, and 500 million medical procedures per year.

[0118] The data analysed in this study correspond to all individual care and medical consumption data that gave rise to reimbursement (consultations and medical devices, hospital admissions, medications and long-term illnesses).

[0119] Statistical methods Classification of sudden death populations Applicants developed a classification model (often referred to as "clustering") of sudden death clusters by utilizing care trajectories observed over the five years prior to cardiac arrest.

[0120] For this purpose, an unsupervised clustering algorithm was used, which allows to identify and describe the heterogeneity of sudden death cases and their associated risk factors.

[0121] More specifically, the algorithm uses language analysis tools and involves the use of a k-means segmentation algorithm.

[0122] Linguistic analysis tools are used to represent patients based on the duration information contained in their care trajectories. In this case, applicants used the Word2Vec algorithm.

[0123] The k-means partitioning algorithm is often referred to as "k-means." This algorithm identifies groups ("clusters") that are predictive of sudden death.

[0124] In the present case, the applicant identified seven clusters.

[0125] First group: The first group is characterized by the following elements: [Table 1]

[0126] [Table 2]

[0127] [Table 3]

[0128] [Table 4]

[0129] Group 2 The second group is characterized by the following elements: [Table 5]

[0130] [Table 6]

[0131] [Table 7]

[0132] [Table 8]

[0133] Group 3 The third group is characterized by the following elements: [Table 9]

[0134] [Table 10]

[0135] [Table 11]

[0136] [Table 12]

[0137] Group 4 The fourth group is characterized by the following elements: [Table 13]

[0138] [Table 14]

[0139] [Table 15]

[0140] [Table 16]

[0141] Group 5 The fifth group is characterized by the following elements: [Table 17]

[0142] [Table 18]

[0143] [Table 19]

[0144] [Table 20]

[0145] Group 6 The sixth group is characterized by the following elements: [Table 21]

[0146] [Table 22]

[0147] [Table 23]

[0148] [Table 24]

[0149] Group 7 The seventh group is characterized by the following elements: [Table 25]

[0150] [Table 26]

[0151] [Table 27]

[0152] [Table 28]

[0153] An algorithm for predicting the occurrence of sudden death Within the framework of these experiments, the Applicant developed an algorithm to predict the occurrence of sudden death within a one-year time frame based on the trajectory of care observed during the five years prior to that event.

[0154] To do this, applicants compared the performance of different supervised statistical classification techniques.

[0155] To do this, Applicants trained and compared each technique using 10-fold linked cross-validation.

[0156] Specifically, the applicants iteratively split the data into two sets, training and testing, in a 9:1 ratio (for every 10 data available, 9 were used for training and 1 for testing), and further modified the sets so that the proportion of sudden deaths (sometimes referred to as SCD, short for Sudden Cardiac Death) in each set was the same.

[0157] Applicants then calculated the area under the curve (AUC), the positive predictive value (PPV), and the sensitivity.

[0158] The three technologies compared are: - First technique T1: Logistic regression, - Second technique T2: Decision trees and - Third technique T3: K nearest neighbors.

[0159] Tables 29 to 32 show the performance of the four models for predicting sudden death after one year.

[0160] The first model M1 corresponds to the SCD comparison with the general population, the second model M2 corresponds to the SCD comparison with acute myocardial infarction (often abbreviated as AMI), the third model M3 corresponds to the SCD comparison with chronic heart failure (often abbreviated as HF), and the fourth model M4 corresponds to the SCD comparison with ischemic heart disease (often abbreviated as IHD).

[0161] The results obtained are as follows: [Table 29]

[0162] [Table 30]

[0163] [Table 31]

[0164] [Table 32]

[0165] Since none of these techniques were satisfactory, applicants turned to neural network techniques.

[0166] Training was performed using the same four previous models on each set of data listed above.

[0167] For this reason, applicants have chosen a recurrent neural network with gates.

[0168] This allows us to obtain significantly improved area under the curve, positive predictive PPV, and sensitivity performance for each of the above models.

[0169] To further improve these results, Applicant used the seven groups determined by the classification technique described above and used a data augmentation technique to artificially increase the number of sudden death cases in each group during the training phase of the neural network, thereby correcting the imbalance in each group (1 sudden death case vs. 12 controls).

[0170] For data augmentation, the applicant used a generative adversarial network. Such networks are often referred to by the acronym GAN, which refers to the corresponding English name of "Generative Adversarial Networks." The applicant has observed in practice that GAN networks are suitable for generating medical data.

[0171] In this way, an algorithm is obtained that predicts the risk of occurrence of sudden death in patients per group. The output of the algorithm developed by the applicant is a quarterly score (so there are four risk scores for a one-year time frame), adapted according to the population (one of seven groups) to which the individual belongs.

[0172] Applicant also used an algorithm to interpret the results. In this case, the algorithm Applicant chose was SAE, an abbreviation for the name "Shapley Additive Explanation," which literally means Shapley Additive Explanation. This provides the most important risk factors that explain these predictions for each individual.

Claims

1. 1. A method for predicting the risk of an occurrence of sudden death in a patient, comprising: The method is computer-implemented, - receiving data regarding the patient's care pathway; determining whether the patient belongs to one of a set of predefined groups from the data relating to the patient's care pathway to obtain determined groups and predefined determined sets of data, each group being associated with a predefined set of data; - for each predetermined determined datum, searching for predetermined datum values ​​of said patient to obtain a set of values ​​specific to said patient; and - applying a neural network to values ​​specific to said patient, said neural network being specific to said determined group, in order to obtain the risk of occurrence of sudden death in said patient; A method including each of the steps.

2. 2. The method of claim 1, wherein each neural network is a recurrent gated neural network.

3. The method of claim 1 , wherein the received data is data regarding the patient's care pathway for the past five years.

4. The method of claim 1 , wherein each predetermined data is the presence of a disorder or the intake of a drug.

5. 2. The method of claim 1, wherein the number of the predetermined data in a group is between 10 and 30.

6. - the group is associated with certain data on respiratory diseases or the intake of respiratory disease-limiting products, - the group is associated with certain data on neurological disorders or the intake of neurologically restricted products, - the group is associated with certain data on cancer disease or the intake of cancer-limiting products, - the group is associated with certain data on addictive disorders or the consumption of addictive disorder-restricting products, - the group is associated with certain data on age-related disorders or the consumption of age-related disorder-limiting products, and - the group is associated with certain data on heart disease or the intake of heart disease-limiting products, The method of claim 1 .

7. 2. The method of claim 1, wherein the groups are obtained by applying a k-means partitioning technique to a set of data comprising data on care pathways for the set of patients, and each predetermined data of a group is obtained by applying a linguistic analysis tool to a set of data comprising data on care pathways for the set of patients.

8. The method of claim 7 , wherein the set of data comprises artificial data generated by applying a function to data relating to care pathways of the set of patients.

9. The method of claim 8 , wherein the function is an adversarial neural network.

10. The prediction method according to claim 1, wherein the number of the predetermined data in the group is between 15 and 25.

11. The method of claim 1, wherein the number of the predetermined data in the group is equal to 20.

12. 12. A computer program product comprising program instructions forming a computer program stored on a readable information medium, said computer program being loadable into a data processing apparatus and which, when run on said data processing apparatus, performs the evaluation method of any one of claims 1 to 11.

13. 12. A readable information medium containing program instructions forming a computer program, the computer program being loadable into a data processing device and which, when run on the data processing device, performs the evaluation method according to any one of claims 1 to 11.

Images