Method and device for stratifying patients for diagnosis, prognosis and response to treatment

Abstract

The invention relates to a method and a device (200) for grouping a group of patients into at least one subgroup of patients on the basis of medical data relating to the patients of the group and of a predetermined predictive task, the predetermined predictive task being associated with at least one disease which affects at least one of the patients of the group. The method and the device make it possible to obtain a model specific to the at least one disease by means of a specialisation method; to create a vector representation of the medical data; and to group patients of the group into at least one subgroup of patients, so as to obtain, for each patient of the group, information on state of health and / or a health result, relating to the at least one disease, based on the grouping of each patient with respect to the at least one subgroup of patients.

Description

METHOD AND DEVICE FOR STRATIFYING SUBJECTS FOR DIAGNOSIS, PROGNOSIS AND RESPONSE TO TREATMENT FIELD OF THE INVENTION

[0001] The present invention relates to the field of artificial intelligence used in precision medicine, for the analysis of medical data. In particular, the invention relates to the development and training of deep learning systems to make predictions of medical outcomes such as a diagnosis or prognosis, to group or stratify subjects, or to discover biomarkers. STATE OF THE ART

[0002] Stratification or grouping of subjects / patients in the medical field consists of identifying groups of subjects with similar characteristics to improve care and treatment effectiveness. This grouping also helps optimize resource management and reduce healthcare costs.

[0003] Defining relevant subject groups is a recurring challenge for medical research. Artificial intelligence has provided new tools to address this issue, which remains a key problem, particularly in the field of immunology and autoimmune diseases. Indeed, these diseases are not yet well understood because their mechanisms of action and modes of expression vary greatly from one subject to another. A relevant grouping of subjects with such diseases would allow for a better understanding of these diseases and therefore better management.

[0004] For this, some studies use unsupervised models to group subjects into subgroups, but this type of approach does not guarantee the relevance of the subgroups. groups obtained for clinical or pharmaceutical trials, since these subgroups are not linked to clinical outcomes.

[0005] Other studies start from clinical results to build a prediction system and then find the predictors that made it possible to better divide the subjects into subgroups by interpreting the mechanisms of action of the prediction system.

[0006] Using a predictive system to extract subgroups ensures that the subgroups are clinically relevant, but requires the use of simple algorithms with limited predictive power for two reasons. The first is that the algorithm used must be explainable to easily extract subgroups. The second is that patient / subject cohorts or clinical trials are often relatively small, especially in autoimmunity, making it impractical to use a complex predictive system that requires extensive data to train. Thus, these two obstacles limit the predictive power of the algorithms and, consequently, the relevance and accuracy of the derived subgroups.

[0007] One of the aims of the present invention is to provide a method and a device for grouping subjects, also called stratification, into homogeneous groups, which overcome the problems previously mentioned while facilitating the diagnosis of certain pathologies, the prognosis, namely the evolution or even the outcome of diseases, or even the identification of personalized treatments and new therapeutic targets by predicting responses to these treatments. SUMMARY

[0008] The invention relates to a device for grouping a group of subjects or patients into at least one subgroup of subjects from medical data relating to said subjects of the group and a predetermined predictive task, said predetermined predictive task being associated with at least one pathology from which at least one of the subjects of the group is affected.

[0009] It should be noted that, within the scope of this description, the terms “patient” and “subject” are used interchangeably. Therefore, any reference to a “patient” should be understood as also including a “subject” and vice versa.

[0010] According to a first embodiment, the device comprises at least one processor and at least one output.

[0011] The at least one processor is configured to: receive a pre-trained foundation model, said pre-trained foundation model having been previously obtained by training, by self-supervised learning, a deep learning model, on a generic unlabeled dataset, the pre-trained foundation model being configured to, from the generic unlabeled dataset, create a generic vector representation; receive a specific labeled dataset comprising the medical data of a plurality of patients, where said specific labeled dataset is labeled according to said predetermined predictive task; obtain a model specific to said at least one pathology by a specialization method via transfer learning, said specialization method consisting of: • training said pre-trained foundation model on said specific dataset; or • adapting said pre-trained foundation model by adding an adapter to said pre-trained foundation model, said adapter consisting of a first neural network trained on said specific data set; receiving medical data relating to the group of patients to be grouped into at least one subgroup; creating a vector representation of the medical data relating to the group of patients received using an algorithm derived from the model specific to said at least one pathology; and grouping patients of the group into at least one subgroup of patients according to said vector representation, so as to obtain, for each patient of the group, information on a health condition and / or a health result, relating to said at least one pathology, according to the grouping of each patient in relation to said at least one subgroup of patients; and

[0012] The at least one output is configured to provide at least one of: the specific model, the vector representation of the medical data relating to the group of patients, the at least one subgroup of patients, said information on the health status and / or the health result relating to said at least one pathology.

[0013] According to a second embodiment, the device comprises at least one processor configured to: receive a pre-trained foundation model, said pre-trained foundation model having been previously obtained by training, by self-supervised learning, a foundation model of the deep learning model type, on a set of unlabeled generic data; receive a set of labeled specific data comprising the medical data of a plurality of subjects, where said set of labeled specific data is labeled according to said predetermined predictive task;obtaining a model specific to said at least one pathology (called “pathology-specific model” in English) by a specialization process via transfer learning (called “transfer learning” in English), said specialization process consisting of: o training said pre-trained foundation model on the specific data set (operation called “fine-tuning” in English) or o adapting said pre-trained foundation model by adding an adapter (called “adapter” or “adaptor” in English) to said pre-trained foundation model, said adapter consisting of a first neural network trained on said specific data set; receiving medical data relating to the group of subjects to be grouped into at least one subgroup; creating a vector representation of the medical data relating to the group of subjects received; and grouping subjects of the group into at least one subgroup of subjects according to said vector representation, so as to obtain, for each subject of the group, information on a state of health and / or a health result, relating to said at least one pathology, according to the grouping of each subject in relation to said at least one subgroup of subjects.

[0014] Obtaining a pathology-specific model by transfer learning from a foundation model pre-trained on a generic, unlabeled dataset has several advantages. On the one hand, it shortens the training time, thanks to the fact that the foundation model has already been pre-trained and to the knowledge acquired during this pre-training. On the other hand, it allows for better performance on a reduced number of training data, knowing that it is very difficult, especially in the field of immunology and immune-mediated inflammatory diseases, to have data and information on subjects or results of clinical trials.

[0015] Obtaining a pathology-specific model by adding an adapter also has advantages, including not having to retrain the pre-trained foundation model (its layers being frozen). Indeed, the adapter corresponds to an overlay of neurons which, once trained, will adapt the pre-trained foundation model to said predetermined predictive task. Only this overlay of neurons needs to be trained on the specific labeled data set. This thus allows for efficiency gains since retraining the foundation model is long and expensive. In addition, it is possible to add as many different adapters as there are different predetermined predictive tasks, having trained each of the adapters separately, without any impact on the pre-trained foundation model. This solution is therefore completely modular as well as simple, fast and economical to implement.

[0016] Generally speaking, the generic vector representation created by the pre-trained foundation model, from the unlabeled generic dataset, allows capturing the important characteristics of this data in a space of reduced dimensionality in order to be able to visualize them, perform exploratory analyses and begin to group similar instances together in order to identify natural groups or classes within this data to facilitate the analysis of large quantities of unstructured data. Thus, an initial grouping of subjects or patients could already be carried out with the pre-trained foundation model but without being specific to the pathology that one wishes to study.

[0017] A major advantage of the invention lies in the vector representation which is created using an algorithm derived from the model specific to said at least one pathology. This vector representation makes it possible to easily and specifically identify subjects having common points relating to the pathology(ies) studied, thanks to the training carried out on the specific data set labeled according to the predetermined predictive task and relating to the pathology(ies) studied, and thus to group them into one or more subgroups. This vector representation also makes it possible to reduce the dimension of the space of the medical data relating to said subjects of the group received as input, which makes it possible to find subgroups more easily.

[0018] Thus, the present device allows, for a given pathology, to group subjects into subgroups in an efficient, relevant and explainable manner, without requiring a significant quantity of training data from clinical trial results.

[0019] In the case of immune-mediated inflammatory diseases or pathologies or any other pathology for which scientific knowledge is limited and / or medical management is complex, this relevant grouping system of subjects suffering from such a disease / pathology can provide a great deal of information such as keys to understanding the mechanisms of action, the identification of specific symptoms or biomarkers, as well as the identification of treatments that seem to work on a certain category of subjects. Indeed, immuno-inflammatory diseases are very heterogeneous, and patients can develop very different forms of the same disease. This heterogeneity is an obstacle to the development of treatments because they rarely work on all the phenotypes of a disease. Stratify the disease by grouping similar patients together would allow treatments to be tested in clinical trials on homogeneous groups of patients and increase the probability of success.

[0020] Thus, for a subject awaiting a diagnosis or treatment for the pathology relating to said predetermined task, part of his medical data or his complete medical file can be injected as input to the model specific to said at least one pathology, which will make it possible to classify the subject studied in a subgroup. Then, depending on this subgroup and the vector representation, to go back to common information such as the treatments which were applied to the subjects of said subgroup, the treatments which worked or not on these subjects of said subgroup, the evolution or prognosis, the identification of one or more biomarkers, etc.

[0021] The medical data may comprise at least two different predetermined modalities. In this case, the processor is also configured to, for each predetermined modality, encode the medical data of the predetermined modality considered separately from the medical data having another modality, this encoding being carried out by an encoding model specific to the predetermined modality considered (called “modality-specific model” in English), said encoding model specific to the predetermined modality considered having been previously trained in a self-supervised manner.

[0022] The medical records of subjects or patients often include data with very different modalities (i.e. formats) such as: free text as in medical reports, tables of digital data such as laboratory results, images from different medical imaging devices (e.g. x-rays, scanners, MRIs, ultrasounds, etc.), or even omics data (e.g. DNA sequences). The advantage of encoding each modality specifically makes it possible to manage several different types of information and therefore to enrich and improve the relevance of grouping subjects into subgroups. Most neural networks and deep learning systems only work with a single type of input data such as images only or just text. Finding a common format and encoding each modality / different type of data in a specific way is thus a way to circumvent the problem and make the most of the performance of the model specific to the pathology(ies) studied.

[0023] In the case where the medical data comprises at least two different predetermined modalities and where the medical data of each predetermined modality is encoded separately from the medical data having another modality, it may be advantageous to merge said encoded medical data using a modality fusion model, such as a second neural network with an attention mechanism for example (called "attention-based neural network" in English). The fusion of the encoded data makes it possible to discover or preserve links / relationships between the medical data of different modalities such as, for example, between an image or a scan and the written medical report associated with it.

[0024] In an advantageous embodiment, the creation of said vector representation of the medical data uses an algorithm derived from the model specific to said at least one pathology, and / or the vector representation of the medical data is representative of contribution information included in the medical data used (or used) to group the subjects in said at least one subgroup. It is this contribution information that will make it possible to better understand the mechanisms of the pathology(ies) studied and to be able to make a better diagnosis, a prognosis or even the identification of personalized treatments and new therapeutic targets for this (these) pathology(ies).

[0025] For example, the algorithm derived from the model specific to said at least one pathology may be an explainability algorithm using artificial intelligence concepts (called "Artificial Intelligence - AI - explainability algorithm" in English), which will allow for an explainable vector representation. That is to say, the vector representation obtained will encode not only the patient's medical information, but also the way in which they contributed interdependently to the final prediction of the task specific to the pathology considered. The subgroups obtained will therefore be more homogeneous, and easier to interpret because they are derived from an algorithm explaining the operation of the specific predictive model. Thus, this will make it easier to make links between the medical data of the subjects of the group and the reasons which led the model specific to said at least one pathology to group the subjects in a certain way into at least one subgroup.

[0026] In an advantageous embodiment, at least one processor is also configured to, after grouping the subjects into at least one subgroup, identify at least one piece of information representative of the interpretation of the grouping by creating a subgroup membership prediction model making it possible to predict membership, for each of the subjects in the group, to said at least one subgroup, said membership prediction model using at least one of: a statistical test, a rules model or a decision tree model, previously trained in an unsupervised manner.

[0027] In particular, said at least one piece of information representative of the interpretation of the grouping is at least one biomarker associated with said subgroup of subjects. The identification of such a biomarker for the pathology(ies) studied will make it possible to better diagnose this / these pathology(ies) and to improve the management of the affected subjects, whether by better monitoring of its / their evolution, the development of specific treatments or the prediction of response to a treatment.

[0028] According to an advantageous embodiment, each predetermined modality is one of the following: electronic health records, clinical notes, laboratory results, medical images, omics data, histopathological data.

[0029] According to an advantageous embodiment, the predetermined predictive task is one of the following: prediction of diagnosis of a disease, prediction of the prognosis of a subject and / or prediction of response to a treatment.

[0030] In order to ensure the relevance of the subgroups and also the proper functioning of the device, the device (via the at least one processor) may also include a model validation process specific to said at least one pathology.

[0031] In the context of the use and application of this device in the field of immunology, it is advantageous that the specific data set includes the medical data of subjects who have been diagnosed or suspected of having an immune-mediated inflammatory disease, and that said at least one pathology with which the predetermined predictive task is associated is an immune-mediated inflammatory disease.

[0032] According to an advantageous embodiment, said at least one pathology is an immune-mediated inflammatory disease, and the predetermined predictive task comprises the prediction of an activity score of said disease, which is representative of the prediction of the prognosis of a subject.

[0033] According to an alternative aspect of the invention, a device for grouping a group of subjects into at least one subgroup of subjects from medical data relating to said subjects of the group and a predetermined predictive task, said predetermined predictive task being associated with at least one pathology from which at least one of the subjects of the group is affected, the device comprising at least one processor configured to: receive an untrained deep learning model type foundation model and an unlabeled generic data set; pre-train, by self-supervised learning, the received untrained foundation model on said unlabeled generic data set; receive a labeled specific data set comprising the medical data of a plurality of subjects, where said labeled specific data set is labeled according to said predetermined predictive task;obtaining a model specific to said at least one pathology by a transfer learning specialization method, said specialization method comprising: o training the pre-trained foundation model on a labeled specific data set (an operation called “fine-tuning” in English); or o adapting said pre-trained foundation model by adding an adapter to said pre-trained foundation model, said adapter consisting of a first neural network trained on said specific data set; receiving medical data relating to the group of subjects to be grouped into at least one subgroup; creating a vector representation of the medical data relating to the group of subjects received; and grouping subjects of the group into at least one subgroup of subjects according to said vector representation, so as to obtain, for each subject of the group, information on a state of health and / or a health result, relating to said at least one pathology, according to the grouping of each subject in relation to said at least one subgroup of subjects.

[0034] According to a second aspect, the invention relates to a computer-implemented method for grouping a group of subjects into at least one subgroup of subjects from medical data relating to said subjects of the group and a predetermined predictive task, said predetermined predictive task being associated with at least one pathology from which at least one of the subjects of the group is affected.

[0035] According to a first embodiment, the method comprises: receiving a pre-trained foundation model, said pre-trained foundation model having been previously obtained by training, by self-supervised learning, a deep learning model, on a set of unlabeled generic data, the pre-trained foundation model being configured to, from the set of unlabeled generic data, create a generic vector representation; receiving a set of labeled specific data comprising the medical data of a plurality of patients, where said set of labeled specific data is labeled according to said predetermined predictive task; obtaining a model specific to said at least one pathology by a specialization method via transfer learning, said specialization method comprising: • training said pre-trained foundation model on said specific labeled dataset; or • adapting said pre-trained foundation model by adding an adapter to said pre-trained foundation model, said adapter consisting of a first neural network trained on said specific data set; receiving medical data relating to the group of patients to be grouped into at least one subgroup; creating a vector representation of the medical data relating to the group of patients received using an algorithm derived from the model specific to said at least one pathology; and grouping the patients of the group into at least one subgroup of patients according to said vector representation, so as to obtain, for each patient of the group, information on a health state and / or a health result, relating to said at least one pathology, according to the grouping of each patient with respect to said at least one subgroup of patients;the provision of at least one of: the specific model, the vector representation of the medical data relating to the group of patients, F at least one subgroup of patients, said information on the state of health and / or the health result relating to said at least one pathology.;

[0036] According to a second embodiment, the method comprises: receiving a pre-trained foundation model, said pre-trained foundation model having been previously obtained by training, by self-supervised learning, a deep learning model type foundation model, on a set of unlabeled generic data; receiving a set of labeled specific data comprising the medical data of a plurality of subjects, where said set of labeled specific data is labeled according to said predetermined predictive task; obtaining a model specific to said at least one pathology by a transfer learning specialization method, said specialization method comprising: o training said pre-trained foundation model on the specific dataset; or o adapting said pre-trained foundation model by adding an adapter to said pre-trained foundation model, said adapter consisting of a first neural network trained on said specific dataset; receiving the medical data relating to the group of subjects to be grouped into at least one subgroup; creating a vector representation of the medical data relating to the group of subjects received; and grouping the subjects of the group into at least one subgroup of subjects according to said vector representation, so as to obtain, for each subject of the group, information on a health state and / or a health result, relating to said at least one pathology, according to the grouping of each subject with respect to said at least one subgroup of subjects.

[0037] According to a third aspect, the invention relates to a computer program product comprising instructions which, when the program is executed by a computer, causes the computer to implement the method for grouping a group of subjects according to any of the previously described embodiments.

[0038] According to a fourth aspect, a computer-readable storage medium comprising instructions that, when the medium is read by a computer, causes the computer to implement the method for grouping a group of subjects according to any of the previously described embodiments. BRIEF DESCRIPTION OF THE FIGURES

[0039] Other objects and advantages of the invention will appear on reading the following description of particular embodiments given by way of illustration, without limitation, the description being made in relation to the appended drawings in which:

[0040] Figure 1 is a representation of steps of a method for grouping a group of subjects into at least one subgroup according to one embodiment from medical data relating to said subjects of the group.

[0041] Figure 2 is a representation of the steps of a method to obtain a pre-trained foundation model.

[0042] Figure 3 is a representation of pre-processing steps for medical data when the data includes several different modalities.

[0043] Figure 4 illustrates a vector representation of the medical data relating to the group of subjects to be grouped into subgroups, and obtained during the method illustrated in Figure 1.

[0044] Figure 5 is a representation of a device for grouping a group of subjects into at least one subgroup according to one embodiment.

[0045] Figure 6 is a schematic representation of the steps and processing carried out on received medical data making it possible to group a group of subjects into at least one subgroup according to one embodiment.

[0046] In the figures, the drawings are not to scale and identical or similar elements are designated by the same references. DETAILED DESCRIPTION

[0047] This description illustrates the principles of this disclosure. It is therefore obvious that persons skilled in the art will be able to devise various arrangements which, although not explicitly described or illustrated herein, embody the principles of the disclosure and are included within its spirit and scope.

[0048] All examples and conditional terms cited in this document are intended to assist the reader in understanding the principles of disclosure and the concepts made by the inventor to the advancement of the art, and must be interpreted as not being limited to the examples and conditions specifically cited.

[0049] Furthermore, all statements of principles, aspects and embodiments of the disclosure, as well as specific examples, are intended to encompass structural and functional equivalents thereof. Furthermore, it is intended that such equivalents include both currently known equivalents and equivalents developed in the future, i.e., any developed element that performs the same function, regardless of structure.

[0050] Thus, for example, those skilled in the art will appreciate that the block diagrams presented herein may represent conceptual views of illustrative circuits incorporating the principles of the disclosure. Similarly, it will be appreciated that all flowcharts, flow diagrams, and the like represent various processes that may be substantially represented on a computer-readable medium and thereby executed by a computer or processor, whether or not that computer or processor is explicitly shown.

[0051] The functions of the various elements illustrated in the figures may be performed by the use of dedicated hardware or hardware capable of executing software in association with appropriate software. When performed by a processor, the functions may be performed by a single dedicated processor, a single shared processor, or several individual processors, some of which may be shared.

[0052] It is to be understood that the elements illustrated in the figures may be implemented in various forms of hardware, software, or combinations thereof. Preferably, these elements are implemented in a combination of hardware and software on one or more suitably programmed general-purpose devices, which may include a processor, memory, and input / output interfaces.

[0053] The present disclosure relates to a computer-implemented method for grouping a group of subjects into at least one subgroup of subjects from medical data relating to said subjects of the group and a predetermined predictive task, said predetermined predictive task being associated with at least one pathology of which at least one at least one of the subjects in the group is affected. The present disclosure also relates to a device configured to implement the method, this device will be detailed after the description of the method.

[0054] The term “subject” refers to an animal, including a human being. For the purposes of this disclosure, a subject may be a patient, namely a person receiving medical care, undergoing or having undergone medical treatment, or being monitored for the development of a disease. The subject group includes at least one subject.

[0055] The term “medical data” refers to any information relating to the health of a subject, namely: electronic medical records (called “Electronic Health Records” or EHR in English) in text format such as written medical reports, medical prescriptions, information on the subject’s sex, age, place of residence and / or professional activity, professional situation, or any other written information that may be related to the subject’s state of health or predispositions such as personal history; information in “table” format, such as analysis results containing tables of numerical values; images from medical imaging devices such as, for example: an ultrasound, an X-ray, a scanner, magnetic resonance imaging (MRI), a tomography, a computed tomography, or histopathology images;“omics” data, such as genomic, transcriptomic, proteomic or metabolomic information, for example;

[0056] These different data types / formats are hereinafter referred to as “modalities”.

[0057] Predetermined predictive task refers to a task that qualified medical personnel could perform or supervise, or have performed or supervised, such as providing a health outcome. For example, but not limited to: establishing a diagnosis or predicting a diagnosis of a disease; the prediction of a subject's prognosis as with the calculation of an activity score (such as that of rheumatoid arthritis - "Disease Activity Score" - DAS 28, or the score relating to Gougerot-Sjôgren's syndrome the "EULAR Sjogren's Syndrom Patient Reported Index" or ESSPRI); and / or the prediction of response to a treatment or the response to a treatment criterion (such as the EULAR response - "European League Against Rheumatism" - in rheumatoid arthritis, or the STAR criterion - "Sjôgre, 's tool for assessing response" - in Gougerot-Sjôgren's syndrome).

[0058] Thus and advantageously, when F at least one pathology is an immune-mediated inflammatory disease, the predetermined predictive task comprises the prediction of an activity score of said disease, which is representative of the prediction of the prognosis of a subject.

[0059] Regarding pathology, this method can be applied to any type of pathology, and regardless of the system (e.g., immune, nervous, respiratory, digestive, cardiovascular, endocrine, reproductive, urinary, musculoskeletal, or integumentary) or organ involved. Preferably, the method applies to immune-mediated inflammatory pathologies or diseases.

[0060] The term "immune-mediated inflammatory diseases / pathologies" refers to diseases characterized by chronic inflammation and an abnormal immune response. These diseases include, but are not limited to: arthritis and rheumatoid arthritis, Crohn's disease, ulcerative colitis, psoriasis, systemic lupus erythematosus, ankylosing spondylitis, celiac disease, Sjögren's syndrome, and Behçet's disease.

[0061] According to a preferred embodiment illustrated in FIG. 1, the method comprises the following main steps which will be described in more detail later: step S 100: receiving a pre-trained foundation model (called “foundation model” in English); step S200: receiving a set of labeled specific data comprising the medical data of a plurality of subjects, wherein said set of labeled specific data is labeled according to said predetermined predictive task; step S300: obtaining a model specific to said at least one pathology (called “pathology-specific model” in English) by a specialization method; step S400: receiving the medical data 10 relating to the group of subjects to be grouped into at least one subgroup 70; step S500: creating a vector representation 60 of the medical data relating to the group of subjects received;and step S600: grouping the subjects of the group into at least one subgroup of subjects 70 according to said vector representation, so as to obtain, for each subject of the group, information on a health state and / or a health result, relating to said at least one pathology, according to the grouping of each subject with respect to said at least one subgroup of subjects 70; step S800: providing at least one output.;

[0062] Step S 100 - reception of the pre-trained foundation model:

[0063] The pre-trained foundation model is a large-scale artificial intelligence model, which is pre-trained on a large amount of generic unlabeled data, including medical data from many subjects / patients with various pathologies. The pre-trained foundation model is a model that can be described as "generalist" because it is suitable for a wide range of downstream tasks.

[0064] Figure 2 illustrates a method for obtaining a pre-trained foundation model. This method comprises three main steps which are: step S10: receiving an untrained foundation model; step S120: receiving an unlabeled generic dataset; and step S150: pre-training the untrained foundation model on the unlabeled generic dataset, which outputs the pre-trained foundation model.

[0065] An untrained foundation model is one that has not yet been exposed to specific data to learn relevant patterns, rules, or characteristics. This means that the model has not yet acquired specific knowledge or capabilities to perform a given task.

[0066] Advantageously, the foundation model is pre-trained to perform the following generic tasks: retrieving information that is masked based on context (also called "masked language modeling"), predicting content for the patient's next visit, predicting a length of hospitalization, predicting the appearance of diseases in the future, for example. In one embodiment, the method of the present invention for clustering a group of subjects into at least one subgroup of subjects 70 comprises the method for obtaining a pre-trained foundation model with the three steps described above.

[0067] It is noted that the pre-training step S 150 may preferably be performed only once, but it may also be repeated on another corresponding training dataset. For example, this may be another larger unlabeled generic dataset to augment the training of the pre-trained foundation model.

[0068] In an alternative embodiment, the method for obtaining a pre-trained foundation model is implemented beforehand, and its result (i.e., the pre-trained foundation model) is stored in a memory. The method of the present invention for grouping a group of subjects into at least one subgroup of subjects 70 is then configured to retrieve / receive this pre-trained foundation model previously generated and stored in the memory.

[0069] Preferably, the foundation model is a deep learning model, more preferably a Transformer. The Transformer works on two main components: attention and feed-forward layers. The Transformer works as follows: Input data encoding: Each input data is encoded into a vector representation and associated with an embedding vector that captures its semantic meaning. Attention: Attention is the key to the Transformer's operation. It calculates the relative importance of different elements of the data received as input to the model compared to each other. In an attention layer, each element of the input data (called a "query") is compared to all other elements (called "keys") to determine their importance. Attention scores measure how relevant elements are to each other. - weighting of input vectors: the attention scores are normalized to obtain attention weights. These weights indicate the relative importance of each element with respect to the query. The embedding vectors of each input data are then weighted by the corresponding attention weights and combined to obtain a contextual representation of each element. anticipation layers: after the attention step, the obtained contextual vectors are transformed by anticipation layers. These layers are generally composed of fully connected neural networks and introduce non-linearities into the model. They make it possible to capture more complex relationships between the elements of the input data. - information propagation: the process of attention and transformation by the anticipation layers is repeated several times, thus forming a stack of layers (usually called a "transformer stack"). This allows for iteration and progressive refinement of the contextual representation of each element of the input data.

[0070] The Transformer has multiple advantages. It can capture long-distance dependencies between elements of the input data. It is capable of performing attention calculations for different elements simultaneously (which speeds up the training process). It easily accepts input data of varying lengths. It offers interpretability thanks to attention mechanisms. It performs well on a wide range of tasks and can be trained on large amounts of data in an unsupervised manner to capture general knowledge before being fine-tuned or adjusted more finely (called "fine-tuning" in English) on a specific task such as the predetermined predictive task.

[0071] Other deep learning models can be used, such as: a model based on recurrent neural networks (RNN), a model based on convolutional neural networks (CNN), a model based on graph neural networks (GNN) or a hybrid model using a combination of transformers, RNN, CNN and / or GNN.

[0072] Step S200 - receipt of the labeled specific data set:

[0073] Data labeling consists of associating a label (or tag) with each data item. “The specific data set is labeled according to the predetermined predictive task” means that the label associated with each data item carries information relating to the predetermined predictive task. For example: in the case where the predetermined predictive task is to predict the response to methotrexate in rheumatoid arthritis, according to the EULAR criteria, it is possible to know, for each subject, whether they have achieved the EULAR response 6 months after the initiation of treatment and therefore to validate the prediction by a dedicated label.

[0074] This data labeling can be done manually or automatically, either because it is contained in the subject's medical file, or using a feed-forward neural network or a machine learning model such as Random Forest or XGboost, for example.

[0075] In the case of the application of the present method for research on immune-mediated inflammatory diseases or their management, it is advantageous that the specific dataset comprises the medical data of subjects who have been diagnosed or suspected of having an immune-mediated inflammatory disease, and where F at least one pathology with which the predetermined predictive task is associated is an immune-mediated inflammatory disease. This dataset specific has been labeled in relation to the diagnosis of immune-mediated inflammatory diseases and / or their management.

[0076] Step S300 - obtaining the model specific to said at least one pathology:

[0077] Obtaining the model specific to said at least one pathology using transfer learning can be done in two different ways. Thus, the specialization method comprises, as desired: training said pre-trained foundation model on said specific data set; or adapting said pre-trained foundation model by adding an adapter to the pre-trained foundation model, said adapter consisting of a first neural network trained on said specific data set.

[0078] In the first case, obtaining the specific model is done by transfer learning from the lessons contained in the specific dataset, by training the pre-trained foundation model in order to refine its weights to specialize it on at least one pathology associated with the predetermined predictive task used for the labeling of the specific dataset. This operation is called "fine-tuning" in English.

[0079] Transfer learning is a machine learning technique that allows the knowledge and skills of a previously trained model (i.e., pre-trained foundation model) to be transferred to a new, similar task, thus facilitating learning on new datasets. In this case, the specialization is relative to the at least one pathology from which at least one of the subjects in the group is affected. Indeed, this specialization is obtained through the training of the pre-trained foundation model carried out with the specific dataset labeled according to the predetermined predictive task, which is itself relative to the at least one pathology.This approach leverages prior knowledge and representations learned by the initial foundation model, leading to improved performance and a reduction in the time and resources required to train a new specific model from scratch.

[0080] In the second case, the use of an adapter makes it possible to avoid having to retrain the pre-trained foundation model, the layers of which can then be frozen. The first neural network, which constitutes the adapter, corresponds to a sort of overlay of neurons that is added to the pre-trained transformation model, which will modify its operation by adapting it to the predetermined predictive task, thanks to the training of the adapter on said specific data set. In addition, this operating mode is very flexible because, for each predetermined predictive task, it is possible to add a corresponding adapter, where each of the adapters has been previously trained separately on a data set specific to the corresponding predetermined predictive task, and this without any impact on the pre-trained foundation model.An adapter can then be added (respectively activated) or removed (respectively deactivated) easily and quickly depending on the needs, the pathology(ies) studied, the medical data available for the group of subjects.

[0081] The integration of the adapter can be done in different ways: by adding the first neural network at the output of the pre-trained foundation model, or inside the pre-trained foundation model by adding new intermediate layers between its different layers. One way to create an adapter is to use the LO RA (Low-Rank Adaptation of Large Language Models) method which freezes the weights of the pre-trained foundation model and injects trainable rank decomposition matrices between each of the layers of the pre-trained foundation model.

[0082] In another advantageous embodiment compatible with the two cases cited, and in order to improve and especially to validate the performance of the model specific to said at least one pathology with respect to the predetermined predictive task for which it is trained, the method comprises a process for validating this model. For example, this validation may be a k-fold cross-validation, where k is an integer greater than or equal to 1, preferably greater than or equal to 5. This validation mode consists of setting aside part of the specific labeled data during training to then be used to evaluate the final model obtained, and this in an iterative manner. One or more cost functions may also be used and minimized during training. It is also possible to use Monte Carlo or “boostrap” cross-validation.

[0083] It should be noted that the training of the pre-trained foundation model to obtain the model specific to said at least one pathology must be carried out specifically for each predetermined predictive task. Consequently, this step is carried out for each specific labeled dataset, in particular relating to different pathologies. Alternatively, this training can also be repeated with another specific labeled dataset, which could be labeled according to a new predetermined predictive task relating to the same pathology as that from which at least one of the subjects of the group is affected (or relating to another pathology), so as to further specialize the specific model on the same pathology (or specialize it on two different pathologies).

[0084] The model specific to said at least one pathology obtained at the end of step S300 makes it possible, according to the predetermined predictive task, to make a prediction as to the state of health and / or the health result of a subject, based on the medical data received and relating to said subject.

[0085] Step S400 - receipt of medical data 10 relating to the group of subjects:

[0086] As stated previously, medical data 10 generally includes different modalities (or formats) ranging from free text such as medical reports, to tables of digital data such as laboratory results, images from different medical imaging devices (e.g., x-rays, scanners, MRIs, ultrasounds, etc.), or even omics data (e.g., DNA sequences) or histopathological data. It may then be advantageous to preprocess them by separately encoding the medical data 10 of different modalities.

[0087] By way of illustration and according to a particular embodiment shown in Figure 2, when the unlabeled generic data set comprises medical data 10 of at least two different predetermined modalities, for each predetermined modality, the medical data of the predetermined modality can be encoded separately from the medical data having another modality corresponding to step S130. This encoding step is optional and can be applied in the same way to each of the data sets received during the steps of the method for grouping the subjects of a group, namely: after receiving the set of specific labeled data and after receiving the medical data 10 of the group of subjects. Thus, if this encoding is applied to one of the data sets used in the method, it must be applied to the other data sets in order to process the data in a similar manner and above all in compatible formats.

[0088] As illustrated in Figure 3, this encoding is performed by a model specific to each predetermined modality considered. Thus, the medical data 10 of modality 1 are encoded by encoder 1, the medical data 10 of modality 2 are encoded by encoder 2 and so on up to N, where N is an integer representing the number of different modalities included in the data set considered. Preferably, these encodings are performed by models that have been previously trained in a self-supervised manner.

[0089] For example: modality 1 can concern medical data in free text format, modality 2 can concern medical data in table format such as laboratory results / analyses, modality 3 can concern medical images in general (i.e. regardless of the type of image and / or the medical device that produced the image such as an MRI image, an ultrasound, a scan, etc.), and modality 4 can concern omics data.

[0090] In an advantageous embodiment, a single encoder can be used to encode information in free text format (such as medical records) and structured laboratory results / analyses whose formats are relatively close. For this, the laboratory results, such as the data tables, are converted into free text format (therefore in the same modality) before being encoded with the medical record data, themselves in free text format, by a large language model (called "Large Language Model" in English) which can include a transformer-type architecture whose operation and technical advantages have been described previously.

[0091] Image data can be encoded by a model with a visual transformer architecture (called "Visual Transformer - ViT" in English). To do this, each image is divided into a grid of regions called "patches". Each patch is then flattened into a vector and transformed into a sequence of input vectors. For example, radiographic images can be resampled or resized to match the size of the images / patches used during (pre)training of the specific image encoding model (e.g.: 224 x 224). In the same way, histopathological images can be divided into sub-images / patches of the same size (e.g.: 224 x 224).

[0092] Advantageously, each input vector is enriched with position information to capture spatial relationships. This can be achieved by adding absolute or relative spatial coordinates to each vector. These enriched input vectors are then fed into multiple attention layers as described previously, with importance weights (or attention scores) calculated for each pair of vectors based on their relationships. This allows, among other things, to focus on parts of the image that are important for the task at hand.

[0093] Omics / genetic data can be encoded via a convolutional neural network.

[0094] It should be noted that one or more modalities can be removed depending on the data that are available for the different training sessions as well as in the medical data 10 of the group of subjects to be grouped. The so-called "predetermined" modalities then correspond to the selected / preserved modalities, and can be: electronic health records, clinical notes, laboratory results, medical images, omics data and / or histopathological data.

[0095] In another advantageous embodiment, the encoded medical data are then merged by a modality fusion model, see step S140 in FIG. 2 in the case of the method for obtaining a pre-trained foundation model. As with encoding, this step of merging the encoded data is optional and can be applied in the same way to each of the encoded data sets during the Steps of the method for grouping subjects in a group. This fusion of encoded data makes it possible to discover or preserve links / relationships between medical data from different modalities. For this, this fusion can be carried out by a second neural network with an attention mechanism that will average the vector representations of the encoded medical data.

[0096] Step S500 - creation of the vector representation:

[0097] As explained previously, the Transformer type foundation model generates a vector representation of the received medical data 10. With the various training sessions carried out as well as the encoding and / or fusion of the data, this vector representation evolves and is refined.

[0098] In an advantageous embodiment, this creation of the vector representation of the medical data 60 uses an algorithm derived from the specific model 20 for said at least one pathology, and where the vector representation 60 of the medical data 10 is representative of contribution information included in the medical data 10 which can be used to group the subjects into said at least one subgroup 70. In one embodiment, the algorithm derived from the specific model 20 is the specific model 20 itself.

[0099] In order to increase the explainability of the method and therefore the explainability of the grouping of subjects into at least one subgroup 70, the algorithm derived from the specific model 20 for said at least one pathology is an explainability algorithm by concepts 40 of artificial intelligence.

[0100] Generally speaking, a concept-based explainability algorithm takes as input a neural network trained for a task (here the specific model 20), and extracts "concepts" (called "features") which are used by the specific model 20 to make its predictions. For each prediction, each concept has a weight (positive or negative) which makes it possible to measure the influence of the concept on the prediction. It is therefore possible to represent each patient by the set of their weights for each concept, which gives a vector which contains the information of the prediction and what influenced it. Working in this space of vectors ensures that the information contained are relevant for the final prediction, unlike the starting space (i.e. the medical data space 10) which may contain information that is useless for the prediction. By finding subgroups 70 in this interpretable space, it is possible to find similar patients for the same prediction. By establishing simple rules between the input variables and the subgroups 70, it is thus possible to obtain biomarkers that characterize each subgroup 70.

[0101] Figure 6 better illustrates this principle and these different sub-steps. Medical data 10 are received and sent into the specific model 20 which generates a first vector representation 'VI' 30 of this data.

[0102] This first vector representation 'VI' 30 is then used as input data by an explainability algorithm 40 which will extract an integer number n of concepts 50, each concept 50 comprising a weight representative of the influence of the concept on the prediction (i.e., an explainability algorithm by artificial intelligence concepts). From these extracted concepts 50, it is then possible to generate a second vector representation 'V2' 60. It is this second vector representation 'V2' 60, created using the algorithm derived from the specific model 20 (i.e., explainability algorithm 40), which is used to group the subjects into at least one subgroup 70, the interpretation of which will be greatly improved compared to the grouping which could have been carried out from the first vector representation 'VI' 30.

[0103] Depending on the medical data 10 used as input, the artificial intelligence concept explainability algorithm 40 used to extract the concepts can be: the SHAP algorithm for "Shapley Additive exPlanations" which is most suitable when the input data is of table type. This algorithm assigns to each input variable a so-called "Shapley" value measuring the impact of this variable on the final prediction, namely the strength of the impact as well as the direction (positive or negative). In this example the "concepts" are therefore exactly the input variables, but they can also be more complex in the context of other algorithms; a sparse autoencoder (SAE) to extract concepts from the latent space of the neural network specific to the predetermined predictive task by training it on one of the neural network's activation layers. The concepts are then the vectors of the decoding matrix of F autoencoder; a matrix factorization algorithm such as independent component analysis (ICA) or non-negative matrix factorization (NMF) applied to one of the neural network's activation layers, to extract concepts and their associated weights.

[0104] According to another embodiment, the artificial intelligence concept explainability algorithm 40 used to extract the concepts can be: AtMan which allows you to obtain explanations on Transformer type models; or Grad CAM which is most suitable when the input data are images. It should be noted that a combination of these algorithms is possible when the input data are of different modalities (i.e. type / format).

[0105] These concept-based explainability algorithms make it possible to know the level of contributions of the different medical data 10 which made it possible to produce the final prediction at the output of the specific model 20 for said at least one pathology.

[0106] Figure 4 illustrates a vector representation of the medical data 10 relating to the group of subjects to be grouped into subgroups 70, and obtained at the output of the specific model 20 for said at least one pathology.

[0107] Step S600 - grouping of group subjects into sub-:

[0108] From the vector representation 60 of the medical data 10, it is easy to group the subjects of the group into at least one subgroup 70 using, for example, distance measures between vectors. Indeed, thanks to the present method, the subjects whose predictions are explained in the same way have close vector representations.

[0109] In the non-limiting example of Figure 4, it is possible to identify 3 possible subgroups 70 (called “clusters” in English): Cl, C2 and C3.

[0110] Step S700 - Identification of at least one piece of information representative of the interpretation [OR I] As illustrated in Figure 1 in step S700 and according to an advantageous embodiment, the method can comprise, after this grouping step S600, the identification of at least one piece of information representative of the interpretation of the grouping by the creation of a model for predicting membership in a subgroup 70 making it possible to predict membership, for each of the subjects of the group, in said at least one subgroup 70.

[0112] Since the links between the medical data 10 of the subjects of the group received as input and the subgroups 70 formed can be complex, one way to simplify and more easily find the information representative of the interpretation of the grouping is to use at least one of: a statistical test, a rule model such as the “skope-rules” (acronym for “Supervised Ky Fan Ordering Permutation for Rule Extraction”), or a decision tree model whose branch level can be of the order of 2, as a membership prediction model and which would be previously trained in an unsupervised manner.

[0113] Preferably, said at least one piece of information representative of the interpretation of the grouping is at least one biomarker associated with said subgroup of subjects 70.

[0114] The term "biomarker" here refers to at least one characteristic that would be objectively measured as an indicator of health, disease, or response to exposure to an intervention, including therapeutic interventions such as the application of a treatment. There are different types of biomarkers. Non-exhaustively, we distinguish between molecular, histological, radiographic, and physiological biomarkers. The BEST glossary (for "Biomarkers, Endpoints and other Tools") defines seven categories of biomarkers: susceptibility / risk, diagnosis, monitoring, prognosis, prediction, pharmacodynamics / response and safety. Finally, the complete identification of a biomarker includes the name of the biomarker, the source / matrix, the measurable characteristic(s) as well as the analytical method used to measure the biomarker which can be a single characteristic or a panel of multiple characteristics.

[0115] In the case of the application of this method for research on immune-mediated inflammatory diseases, the identification of biomarkers will make it possible to improve diagnosis, prognosis but also management by applying a treatment adapted to the affected subject.

[0116] Step S 800: Provision of at least one output

[0117] This step consists of providing at least one of: the specific model 20, the vector representation 60 of the medical data relating to the group of patients, F at least one subgroup of patients 70, said information on the state of health and / or the health result relating to said at least one pathology. Each of these outputs has its advantages: the specific model 20 can subsequently be used to classify a group of subjects according to their medical data received as input; the vector representation 60 of the medical data 10 relating to the group of patients will allow a better visualization of the distribution of certain factors / certain particular characteristics that the specific model 20 has “pre-classified” in order in particular to find specific biomarkers; F at least one subgroup of patients 70 can, among other things, allow better care for these people or even to carry out a clinical study on a drug used (or tested) by this subgroup 70, as well as information on the state of health and / or the health result relating to said at least one pathology.

[0118] The method for grouping a group of subjects into at least one subgroup 70, as previously described in the various embodiments, can be implemented in a device 200 as illustrated in FIG. 5.

[0119] According to one embodiment, the device 200 comprises at least one processor 202 configured to carry out the different steps of the method previously described, such as: step S 100: receiving a pre-trained foundation model, said pre-trained foundation model having been previously obtained by training, by self-supervised learning, a Transformer type foundation model, on a set of unlabeled generic data; step S200: receiving a set of labeled specific data comprising the medical data of a plurality of subjects, where said set of labeled specific data is labeled according to said predetermined predictive task; step S300: obtaining a model specific to said at least one pathology by training said pre-trained foundation model on said set of labeled specific data using transfer learning;step S400: receiving medical data 10 relating to the group of subjects to be grouped into at least one subgroup 70; step S500: creating a vector representation 60 of the medical data relating to the group of subjects received; and step S600: grouping subjects of the group into at least one subgroup of subjects 70 according to said vector representation, so as to obtain, for each subject of the group, information on a state of health and / or a health result, relating to said at least one pathology, according to the grouping of each subject with respect to said at least one subgroup of subjects 70.;

[0120] The at least one processor 202 may also be configured to perform the optional step S700 previously described.

[0121] Advantageously, the processor also comprises a generation module, configured to implement the method for obtaining a pre-trained foundation model as described previously.

[0122] In a particular embodiment, the processor comprises a data processing module configured to encode and / or merge the medical data. when these include at least two different predetermined modalities, as described above. Advantageously, this data processing module is also capable of selecting (or deleting) predetermined modalities that one wishes to keep and process (reciprocally that one wishes to delete and not process).

[0123] In another advantageous embodiment, the processor comprises an interpretation module configured to identify at least one piece of information representative of the interpretation of the grouping carried out. For this, the interpretation module is capable of creating a subgroup membership prediction model 70 making it possible to predict the membership, for each of the subjects of the group, of said at least one subgroup 70, said membership prediction model using at least one of: a statistical test, a rules model or a decision tree model, previously trained in an unsupervised manner.

[0124] According to another embodiment, the device 200 also comprises a memory 201 for storing program instructions loadable into the processor 202 such as the execution of the different steps of the method previously described and illustrated in Figures 1 and 2. The memory 201 can also store data and information useful for carrying out the steps of the method as described above.

[0125] The processor 202 may be, for example: a processor or a processing unit capable of interpreting instructions in a computer language, the processor or the processing unit being able to comprise, be associated with or be attached to a memory comprising the instructions, or the association of a processor / processing unit and a memory, the processor or the processing unit capable of interpreting instructions in a computer language, the memory comprising said instructions, or an electronic card in which the steps of the invention are described within the silicon, or a programmable electronic chip such as an FPGA chip (for "Field-Programmable Gate Array").

[0126] The device 200 may also comprise an input interface 203 for receiving input data and an output interface 204 for providing output data (i.e. configured to perform the step S800 previously described). Examples of input data may be the medical data relating to the subjects of the group or the different training data sets. Examples of output data may be the subgroup distribution of each subject 70, the data for interpreting the formation of these subgroups 70 such as the information representative of the interpretation of the grouping, or the vector representation of the medical data.

[0127] To facilitate interaction with the device 200, a display 205 and a keyboard 206 may be provided and connected to the processor / computer circuit 202.

[0128] The present disclosure also relates to a computer program product for providing the optimal sequence of actions causing a complex system to evolve from an initial state to a final state, the computer program product comprising instructions which, when the program is executed by a computer, cause the computer to automatically perform the steps of the method according to one of the embodiments described above.

[0129] The computer program product for performing the method described above may be written in the form of computer programs, code segments, instructions, or any combination thereof, to individually or collectively instruct or configure the processor or computer to operate as a special-purpose machine or computer to perform the operations performed by the hardware components. In one example, the computer program includes machine code directly executed by a processor or computer, such as machine code produced by a compiler. In another example, the computer program product includes higher-level code that is executed by a processor or computer using an interpreter. Programmers of ordinary skill in the art can readily write the instructions or software based on the block diagrams and flowcharts illustrated in the drawings and corresponding descriptions in the specification, which disclose algorithms for performing the operations of the method as described above.

[0130] The present disclosure also relates to a computer-readable storage medium comprising instructions which, when the program is executed by a computer, cause the computer to perform the steps of the method according to one of the embodiments described above.

[0131] According to one embodiment, the computer-readable storage medium is a non-transitory computer-readable storage medium.

[0132] Computer programs implementing the method described in the present embodiments may generally be distributed to users on a computer-readable storage medium, such as, but not limited to, an SD card, an external storage device, a microchip, a flash memory device, a portable hard drive, and software websites. From the distribution medium, the computer programs may be copied to a hard drive or similar intermediate storage medium. The computer programs may be executed by loading the computer instructions, either from their distribution medium or from their intermediate storage medium, into the execution memory of the computer, configuring the computer to act in accordance with the method of the present disclosure.All these operations are well known to people skilled in computer systems.

[0133] The instructions or software for controlling a processor or computer to implement the hardware components and perform the methods described above, and all associated data, data files, and data structures, are recorded, stored, or fixed in or on one or more non-transitory computer-readable storage media. Examples of non-transitory computer-readable storage media include read-only memory (ROM), random access memory (RAM), flash memory, CD-ROMs, CD-Rs, CD-I-Rs, CD-RWs, CD-I-RWs, DVD-ROMs, DVD-Rs, DVD+Rs, DVD+Rs, DVD+Rs, DVD+Rs, DVD+Rs, DVD+Rs, DVD+Rs, DVD+Rs, DVD+Rs, DVD+Rs, DVD+Rs, DVD+Rs, DVD+Rs, DVD+Rs, DVD+Rs, DVD+Rs, etc. DVD- Rs, DVD+ Rs, DVD- RWs, DVD+ RWs, DVD- RAMs, BD- ROMs, BD- Rs, BD- R LTHs, BD- Res, magnetic tapes, floppy disks, magneto-optical data storage devices, optical data storage devices, hard disk drives, solid-state drives, and any device known to a person skilled in the art capable of storing the instructions or software and all associated data, data files and data structures in a non-transitory manner and of providing the instructions or software and all associated data, data files and data structures to a processor or computer so that the processor or computer can execute the instructions.In one example, the instructions or software and all associated data, data files, and data structures are distributed across network-coupled computer systems, such that the instructions and software and all associated data, data files, and data structures are stored, accessed, and executed in a distributed manner by the processor or computer.

Claims

CLAIMS 1. A device (200) for grouping a group of patients into at least one subgroup of patients (70) from medical data relating to said patients of the group (10) and a predetermined predictive task, said predetermined predictive task being associated with at least one pathology with which at least one of the patients of the group is affected, the device comprising: at least one processor (202) configured to: a) receive a pre-trained foundation model (S 100), said pre-trained foundation model having been previously obtained by training, by self-supervised learning, a deep learning model, on a set of unlabeled generic data, the pre-trained foundation model being configured to, from the set of unlabeled generic data, create a generic vector representation (30);b) receiving a labeled specific data set comprising the medical data of a plurality of patients, wherein said labeled specific data set is labeled according to said predetermined predictive task (S200); c) obtaining a specific model (20) for said at least one pathology by a specialization method via transfer learning (S300), said specialization method consisting of:; ■ training said pre-trained foundation model on said specific dataset; or ■ adapting said pre-trained foundation model by adding an adapter to said pre-trained foundation model, said adapter consisting of a first neural network trained on said specific data set; d) receiving medical data (10) relating to the group of patients to be grouped into at least one subgroup (S400); e) creating a vector representation (60) of the medical data (10) relating to the group of patients received (S500) using an algorithm (40) derived from the specific model (20) for said at least one pathology; and f) grouping patients of the group into at least one subgroup of patients (70) according to said vector representation (60), so as to obtain, for each patient of the group, information on a health state and / or a health result, relating to said at least one pathology, according to the grouping of each patient with respect to said at least one subgroup of patients (S600); and at least one output configured to provide at least one of: the specific model (20), the vector representation (60) of the medical data (10) relating to the group of patients, F at least one subgroup of patients (70), said information on the health state and / or the health result relating to said at least one pathology.

2. Device (200) according to claim 1, wherein, when the medical data (10) comprises at least two different predetermined modalities, for each predetermined modality, the medical data of the predetermined modality considered are encoded separately from the medical data having another modality, the encoding being carried out by an encoding model specific to the predetermined modality considered, said encoding model specific to the predetermined modality considered having been previously trained in a self-supervised manner (S 130).

3. The device (200) of claim 2, wherein said encoded medical data corresponding to said at least two different predetermined modalities are fused using a modality fusion model (S 140).

4. Device (200) according to claim 3, wherein said modality fusion model is a second neural network with attention mechanism.

5. Device (200) according to any one of the preceding claims, wherein the vector representation (60) of the medical data (10) is representative of contribution information included in the medical data (10) used to group the patients into said at least one subgroup.

6. Device (200) according to claim 5, wherein the algorithm derived from the specific model (20) for said at least one pathology is an algorithm of explainability by artificial intelligence concepts (40).

7. Device (200) according to any one of the preceding claims, in which the at least one processor is also configured to, after grouping the patients into at least one subgroup (S600), identify at least one piece of information representative of the interpretation of the grouping by creating a subgroup membership prediction model making it possible to predict the membership, for each of the patients in the group, of said at least one subgroup, said membership prediction model using at least one of: a statistical test, a rules model or a decision tree model, previously trained in an unsupervised manner (S700).

8. Device (200) according to claim 7 wherein said at least one piece of information representative of the interpretation of the grouping is at least one biomarker associated with said subgroup of patients (70).

9. Device (200) according to one of the preceding claims, wherein each predetermined modality is one of the following: electronic health records, clinical notes, laboratory results, medical images, omics data, histopathological data.

10. Device (200) according to one of the preceding claims, wherein the predetermined predictive task is one of the following: prediction of diagnosis of a disease, prediction of the prognosis of a patient and / or prediction of response to a treatment.

11. Device (200) according to one of the preceding claims, in which the at least one processor also comprises the validation of the specific model (20) for said at least one pathology.

12. Device (200) according to one of the preceding claims, in which the specific data set comprises the medical data of patients having been diagnosed or suspected of having an immune-mediated inflammatory disease, and wherein said at least one pathology with which the predetermined predictive task is associated is an immune-mediated inflammatory disease.

13. Device (200) according to one of the preceding claims, wherein when the at least one pathology is an immune-mediated inflammatory disease, the predetermined predictive task comprises the prediction of an activity score of said disease, which is representative of the prediction of the prognosis of a patient.

14. A computer-implemented method for grouping a group of patients into at least one patient subgroup (70) from medical data (10) relating to said patients of the group and a predetermined predictive task, said predetermined predictive task being associated with at least one pathology with which at least one of the patients of the group is affected, said method comprising: a) receiving a pre-trained foundation model (S 100), said pre-trained foundation model having been previously obtained by training, by self-supervised learning, a deep learning model, on a set of unlabeled generic data, the pre-trained foundation model being configured to, from the set of unlabeled generic data, create a generic vector representation (30);b) receiving a labeled specific data set comprising the medical data of a plurality of patients, wherein said labeled specific data set is labeled according to said predetermined predictive task (S200); c) obtaining a specific model (20) for said at least one pathology by a specialization method via transfer learning (S300), said specialization method comprising:; ■ training said pre-trained foundation model on said specific labeled dataset; or ■ adapting said pre-trained foundation model by adding an adapter to said pre-trained foundation model, said adapter consisting of a first neural network trained on said specific data set; d) receiving the medical data (10) relating to the group of patients to be grouped into at least one subgroup (S400); e) creating a vector representation (60) of the medical data (10) relating to the group of patients received using an algorithm derived from the specific model (20) for said at least one pathology (S500); and f) grouping the patients of the group into at least one subgroup of patients (70) according to said vector representation (60), so as to obtain, for each patient of the group, information on a health state and / or a health result, relating to said at least one pathology, according to the grouping of each patient with respect to said at least one subgroup of patients (S600);g) providing at least one of: the specific model (20), the vector representation (60) of the medical data (10) relating to the group of patients, the at least one subgroup of patients (70), said information on the state of health and / or the health result relating to said at least one pathology.; 15. A computer program product comprising instructions which, when the program is executed by a computer, causes the computer to implement the method for grouping a group of patients according to claim 14.

Images