METHOD AND DEVICE FOR STRATIFYING SUBJECTS FOR DIAGNOSIS, PROGNOSIS AND RESPONSE TO TREATMENT

Abstract

METHOD AND DEVICE FOR STRATIFYING SUBJECTS FOR DIAGNOSIS, PROGNOSIS AND RESPONSE TO TREATMENT The invention relates to a method and a device (200) for grouping a group of subjects into at least one subgroup of subjects from medical data relating to said subjects in 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 in the group is affected.The method and apparatus make it possible to obtain a model specific to at least one pathology through a specialization process; to create a vector representation of medical data; and to group subjects from the group into at least one subgroup of subjects, so as to obtain, for each subject in the group, information on a health status and / or a health outcome, related to at least one pathology, according to the grouping of each subject with respect to said at least one subgroup of subjects. Figure from the abstract: Fig. 1.

Description

Title of the invention: METHOD AND DEVICE FOR STRATIFYING SUBJECTS FOR DIAGNOSIS, PROGNOSIS AND RESPONSE TO TREATMENT FIELD OF 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 (“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 the effectiveness of treatments. This grouping also optimizes resource management and reduces healthcare costs.

[0003] Defining relevant subject groups is a recurring challenge for medical research. Artificial intelligence has provided new tools to try to solve this problem, which remains a key issue, particularly in the field of immunology and autoimmune diseases. Indeed, these diseases are not yet well understood because they have very different mechanisms of action and modes of expression from one individual 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 purpose, some studies use unsupervised models to group subjects in subgroups, but this type of approach does not guarantee the relevance of the subgroups obtained for clinical or pharmaceutical trials, given that 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 which made it possible to better divide the subjects into subgroups by interpreting the mechanisms of action of the prediction system.

[0006] The use of 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 in order to easily extract subgroups. The The second is that patient / subject cohorts or clinical trials are often relatively small, particularly in autoimmunity, making it impossible to use a complex predictive system requiring large datasets for training. 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 mentioned above while facilitating the diagnosis of certain pathologies, the prognosis, namely the evolution or even the outcome of diseases, or 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 into at least one subgroup of subjects based on medical data relating to said subjects in the group and a predetermined predictive task, said predetermined predictive task being associated with at least one pathology affecting at least one of the subjects in the group, the device comprising at least one processor configured to: - receive a pre-trained foundation model, said pre-trained foundation model having been previously obtained by training, through self-supervised learning, a deep learning model type foundation model on a generic unlabeled dataset; - receive a specific labeled dataset comprising the medical data of a plurality of subjects, where said specific labeled dataset is labeled according to said predetermined predictive task; - to obtain a model specific to at least one pathology (called a "pathology-specific model" in English) through a specialization process via transfer learning (called "transfer learning" in English), said specialization process consisting of: • training said pre-trained foundation model on the specific dataset (an operation called "fine-tuning") or • the adaptation of said pre-trained foundation model by adding an adapter (called "adapter" or "adaptor" in English) to said pre-trained foundation model, said adapter being made up of a first neural network trained on said specific dataset; - receive medical data relating to the group of subjects to be grouped in at least one subgroup; - create a vector representation of the medical data relating to the group of subjects received; and - to group subjects from the group into at least one subgroup of subjects according to said vector representation, so as to obtain, for each subject in the group, information on a state of health and / or a health outcome, relating to said at least one pathology, according to the grouping of each subject in relation to said at least one subgroup of subjects.

[0009] Obtaining a pathology-specific model through transfer learning from a foundation model pre-trained on a generic, unlabeled dataset has several advantages. Firstly, it shortens the training time, thanks to the fact that the foundation model has already been pre-trained and the knowledge acquired during this pre-training. Secondly, it allows for better performance with a smaller amount of training data, given the difficulty, especially in the field of immunology and immune-mediated inflammatory diseases, of obtaining data and information on subjects or clinical trial results.

[0010] Obtaining a pathology-specific model by adding an adapter also offers advantages, notably that of 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 the predetermined predictive task. Only this overlay of neurons needs to be trained on the specific labeled dataset. This thus increases efficiency since retraining the foundation model is lengthy and costly. Furthermore, it is possible to add as many different adapters as there are different predetermined predictive tasks, having trained each adapter separately, without any impact on the pre-trained foundation model. This solution is therefore completely modular, as well as simple, quick, and economical to implement.

[0011] Vector representation, for its part, makes it easy to identify subjects with commonalities related to the pathology(ies) under study, thanks to training performed on specific datasets labeled according to the predetermined predictive task and related to the pathology(ies) under study, and thus to group them into one or more subgroups. This vector representation also makes it possible to reduce the dimensionality of the medical data space relating to the subjects in the input group, which facilitates the identification of subgroups.

[0012] Thus, the present device makes it possible, for a given pathology, to group subjects into subgroups in an efficient, relevant and explainable manner, without requires a significant amount of training data derived from clinical trial results.

[0013] In the case of immune-mediated inflammatory diseases or pathologies or any other pathology of which scientific knowledge is limited and / or medical management is complex, this relevant grouping device for subjects with 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.

[0014] Thus, for a subject awaiting diagnosis or treatment for the pathology related to the predetermined task, a portion of their medical data or their complete medical file can be input into the model specific to at least one pathology, allowing the subject to be classified into a subgroup. Then, based on this subgroup and the vector representation, common information can be retrieved, such as the treatments applied to the subjects in said subgroup, whether the treatments were successful or not, the evolution or prognosis, the identification of one or more biomarkers, etc.

[0015] Medical data may include 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 a model specific to the predetermined modality considered (called a "modality-specific model" in English), said model specific to the predetermined modality considered having been previously trained in a self-supervised manner.

[0016] Subjects' or patients' medical records often include data with very different modalities (i.e., formats), such as: free text like in medical reports, tables of numerical data like laboratory results, images from various medical imaging devices (e.g., X-rays, CT scans, MRIs, ultrasounds), or even omics data (e.g., DNA sequences). The advantage of specifically encoding each modality is that it allows for the management of several different types of information and thus enriches and improves the relevance of grouping subjects into subgroups. Most neural networks and deep learning systems work with only one type of input data, such as images only or just text.Finding a common format and encoding each different modality / data type 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.

[0017] In cases where medical data comprise at least two different predetermined modalities and where the medical data for each predetermined modality are encoded separately from the medical data for another modality, it may be advantageous to fuse said encoded medical data using a modal fusion model, such as a second neural network with an attention mechanism (called an "attention-based neural network"). The fusion of encoded data makes it possible to discover or preserve links / relationships between medical data of different modalities, such as, for example, between an image or scan and its associated written medical report.

[0018] In an advantageous embodiment, the creation of said vector representation of medical data uses an algorithm derived from the model specific to said at least one pathology, and the vector representation of medical data is representative of contribution information contained in the medical data used to group the subjects into said at least one subgroup. It is this contribution information that will allow for a better understanding of the mechanisms of the pathology(ies) studied and enable improved diagnosis, prognosis, or the identification of personalized treatments and new therapeutic targets for this / these pathology(ies).

[0019] For example, the algorithm derived from the model specific to said at least one pathology may be an artificial intelligence explainability algorithm (called an "Artificial Intelligence - AI - explainability algorithm" in English), which will allow for an explainable vector representation, that is, for understanding the result obtained, namely why such a vector representation was created and not another. This will also make it easier to link the medical data of the subjects in the group with the reasons that led the model specific to said at least one pathology to group the subjects in a certain way into at least one subgroup.

[0020] In an advantageous embodiment, the device comprises, after grouping the subjects into at least one subgroup, the identification of at least one piece of information representative of the explainability of the grouping by the creation of a subgroup membership prediction model allowing the prediction of membership, for each of the subjects in the group, of said at least one subgroup, said membership prediction model using at least one of: a statistical test, a rule model or a decision tree model, previously trained in an unsupervised manner

[0021] In particular, said at least one piece of information representative of the explainability 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 care of affected subjects, whether by better monitoring of its / their evolution, the development of specific treatments or the prediction of response to a treatment.

[0022] 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.

[0023] According to an advantageous embodiment, the predetermined predictive task is one of the following: prediction of the diagnosis of a disease, prediction of the prognosis of a subject and / or prediction of response to a treatment.

[0024] In order to guarantee the relevance of the subgroups and also the proper functioning of the device, the device may also include a validation process of the model specific to said at least one pathology.

[0025] In the context of the use and application of the present device to the field of immunology, it is advantageous that the specific dataset include 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 to which the predetermined predictive task is associated is an immune-mediated inflammatory disease.

[0026] According to an advantageous embodiment, said at least one pathology is an immune-mediated inflammatory disease, and the predetermined predictive task includes the prediction of an activity score of said disease, which is representative of the prediction of the prognosis of a subject.

[0027] According to an alternative aspect of the invention, a device for grouping a group of subjects into at least one subgroup of subjects based on medical data relating to said subjects in the group and a predetermined predictive task, said predetermined predictive task being associated with at least one pathology affecting at least one of the subjects in the group, the device comprising at least one processor configured to: - receive a foundation model of the type untrained deep learning model and a generic unlabeled dataset; - pre-train, through self-supervised learning, the untrained foundation model received on said unlabeled generic dataset; - receive a specific labeled dataset comprising the medical data of a plurality of subjects, where said specific labeled dataset is labeled according to said predetermined predictive task; - to obtain a model specific to at least one pathology by a specialization process through transfer learning, said specialization process- reading including: • training the pre-trained foundation model on a specific labeled dataset (an operation called "fine-tuning"); or • the adaptation of said pre-trained foundation model by adding an adapter to said pre-trained foundation model, said adapter being made up of a first neural network trained on said specific dataset; - receive medical data relating to the group of subjects to be grouped into at least one subgroup; - create a vector representation of the medical data relating to the group of subjects received; and - to group subjects from the group into at least one subgroup of subjects according to said vector representation, so as to obtain, for each subject in the group, information on a state of health and / or a health outcome, relating to said at least one pathology, according to the grouping of each subject in relation to said at least one subgroup of subjects.

[0028] 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 based on medical data relating to said subjects in 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 in the group is affected, the method comprising: - 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 generic unlabeled dataset; - the receipt of a specific labeled dataset comprising the medical data of a plurality of subjects, where said specific labeled dataset is labeled according to said predetermined predictive task; - obtaining a model specific to at least one pathology through a specialization process using transfer learning (called "transfer learning" in English), said specialization process comprising: • training said pre-trained foundation model on the specific dataset; or • the adaptation of 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; - the receipt of medical data relating to the group of subjects to be grouped into at least one subgroup - the creation of a vector representation of the medical data relating to the group of subjects received; and - the grouping of 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 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.

[0029] 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 one of the embodiments previously described.

[0030] According to a fourth aspect, a computer-readable storage medium comprising instructions which, when the medium is read by a computer, cause the computer to implement the method for grouping a group of subjects according to any of the embodiments previously described. BRIEF DESCRIPTION OF THE FIGURES

[0031] Other objects and advantages of the invention will become apparent from the following description of particular embodiments given by way of illustration, not limitation, the description being made in relation to the accompanying drawings in which:

[0032] [Fig. 1] The [Fig. 1] is a representation of steps of a method for grouping a group of subjects into at least one subgroup according to an embodiment based on medical data relating to said subjects of the group.

[0033] [Fig.2] The [Fig.2] is a representation of the steps of a method for obtaining a pre-trained foundation model.

[0034] [Fig.3] The [Fig.3] is a representation of pre-processing steps of medical data when these include several different modalities.

[0035] [Fig.4] [Fig.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 [Fig.1].

[0036] [Fig.5] The [Fig.5] is a representation of a device for grouping a group of subjects into at least one subgroup according to one embodiment.

[0037] In the figures, the drawings are not to scale and identical or similar elements are designated by the same references. DETAILED DESCRIPTION

[0038] This description illustrates the principles of this disclosure. It is therefore understood that persons competent in this field may devise various arrangements which, although not explicitly described or illustrated in this document, embody the principles of the disclosure and are understood in its spirit and scope.

[0039] All examples and conditional terms cited in this document are intended to assist the reader in understanding the principles of disclosure and the concepts brought by the inventor to the advancement of the art, and should be interpreted as not being limited to the examples and conditions specifically cited.

[0040] Furthermore, all statements of principles, aspects, and achievements of the disclosure, as well as specific examples, are intended to encompass their structural and functional equivalents. Moreover, it is intended that these equivalents will include both currently known equivalents and equivalents developed in the future—that is, any developed element that performs the same function, regardless of structure.

[0041] Thus, for example, those skilled in the art will appreciate that the functional diagrams presented herein may represent conceptual views of illustrative circuits incorporating the principles of disclosure. Similarly, it will be appreciated that all organizational charts, flowcharts, and the like represent various processes that can be substantially represented on a computer-readable medium and thus executed by a computer or processor, whether or not that computer or processor is explicitly shown.

[0042] The functions of the various elements illustrated in the figures can be performed using dedicated hardware or hardware capable of running software in conjunction with suitable software. When performed by a processor, the functions can be carried out by a single dedicated processor, a single shared processor, or several individual processors, some of which may be shared.

[0043] It should be understood that the elements illustrated in the figures can 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 appropriately programmed multipurpose devices, which may include a processor, memory, and input / output interfaces.

[0044] This disclosure relates to a computer-implemented method for grouping a group of subjects into at least one subgroup of subjects based on medical data relating to said subjects in the group and a predetermined predictive task, said predetermined predictive task being associated with at least one pa theology in which at least one of the subjects in the group is affected. This disclosure also concerns a device configured to implement the method; this device will be detailed after the description of the method.

[0045] 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 in the context of the development of a disease. The group of subjects comprises at least one subject.

[0046] The term "medical data" refers to any information relating to the health of a subject, namely: - electronic health 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, employment status, or any other written information that may be related to the subject's state of health or predispositions such as their personal history; - information in "table" format, such as analysis results containing tables of numerical values; - images from medical imaging devices such as, for example: ultrasound, X-ray, CT scan, magnetic resonance imaging (MRI), tomography, computed tomography, or histopathology images; - “omics” data, such as genomic, transcriptomic, proteomic or metabolomic information, for example.

[0047] These different types / formats of data are hereafter referred to as "modalities".

[0048] The 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; - predicting a subject's prognosis, such as by calculating an activity score (like that of rheumatoid arthritis - "Disease Activity Score" - DAS 28, or the score related to Sjögren's syndrome, the "EULAR Sjögren's Syndrome 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 Disease Against Rheumatism" - in rheumatoid arthritis, or the STAR criterion - "Sjögre's tool for assessing response" - in Sjögren's syndrome) Sjögren).

[0049] Thus and advantageously, when at least one pathology is an immune-mediated inflammatory disease, the predetermined predictive task includes the prediction of an activity score of said disease, which is representative of the prediction of the prognosis of a subject.

[0050] Regarding the pathology, the present method can be applied to any type of pathology, regardless of the system (e.g., immune, nervous, respiratory, digestive, cardiovascular, endocrine, reproductive, urinary, musculoskeletal, or integumentary) or organ involved. Preferably, the method is applied to immune-mediated inflammatory pathologies or diseases.

[0051] The terms "immune-mediated inflammatory diseases" refer 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.

[0052] 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 S100: receiving a pre-trained foundation model (called "foundation model" in English); - step S200: the receipt of a specific labeled dataset comprising the medical data of a plurality of subjects, where said specific labeled dataset is labeled according to said predetermined predictive task; - step S300: obtaining a model specific to at least one pathology (called "pathology-specific model" in English) by a specialization process; - step S400: the receipt of medical data relating to the group of subjects to be grouped into at least one subgroup; - Step S500: the creation of a vector representation of the medical data relating to the group of subjects received; and - step S600: the grouping of 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 state of health and / or a health outcome, relating to said at least one pathology, according to the grouping of each subject in relation to said at least one subgroup of subjects.

[0053] Step S100 - Reception of the pre-trained foundation model:

[0054] The pre-trained foundation model is a large-scale artificial intelligence model that is pre-trained on a large amount of generic, unlabeled data, including medical data from numerous subjects / patients with diverse and varied pathologies. The pre-trained foundation model can be described as a "generalist" model because it is suitable for a wide range of downstream tasks.

[0055] Figure 2 illustrates a method for obtaining a pre-trained foundation model. This method comprises three main steps, which are: - SI step 10: receiving an untrained foundation model; - Step S120: Receiving a generic, unlabeled dataset; and - Step S150: Pre-training the untrained foundation model on the unlabeled generic dataset, which outputs the pre-trained foundation model.

[0056] An untrained foundation model is a model that has not yet been exposed to specific data to learn relevant patterns, rules, or features. This means that the model has not yet acquired specific knowledge or capabilities to perform a given task.

[0057] Advantageously, the foundation model is pre-trained to perform the following generic tasks: retrieving information that is masked according to the context (also called "masked language modeling" in English), predicting content for the patient's next visit, predicting a length of hospitalization, predicting the onset of diseases in the future, for example.

[0058] In one embodiment, the method of the present invention for grouping a group of subjects into at least one subgroup of subjects comprises the method for obtaining a pre-trained foundation model with the three steps described above.

[0059] It should be noted that the pre-training step S150 can preferably be performed only once, but it can also be repeated on another corresponding training dataset. For example, this could be another, larger, unlabeled generic dataset to augment the training of the pre-trained foundation model.

[0060] 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 memory. The method of the present invention for grouping a group of subjects into at least one subgroup of subjects is then configured to retrieve / receive this pre-trained foundation model previously generated and stored in memory.

[0061] Preferably, the foundation model is a deep learning model, more Preferably a Transformer. The Transformer's operation relies on two main components: attention and feed-forward layers. The Transformer's operation is as follows: - Encoding of input data: each input data is encoded into a vector representation and associated with an embedding vector that captures its semantic meaning. Attention: Attention is key to the Transformer's operation. It allows the calculation of the relative importance of the different elements of the input data to the model. 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 the elements are to each other. - Weighting of input vectors: attention scores are normalized to obtain attention weights. These weights indicate the relative importance of each element in relation to the query. The nesting vectors of each input data point are then weighted by the corresponding attention weights and combined to obtain a contextual representation of each element. - Anticipation layers: After the attention stage, the resulting 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 allow for the capture of 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 (generally called a "transformer stock"). This allows for the iteration and progressive refinement of the contextual representation of each element of the input data.

[0062] The Transformer has multiple advantages. It allows for the capture of long-range 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 readily accepts input data of varying lengths. It offers a degree of interpretability through its 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 tuned or fine-tuned (called "fine-tuning" in English) on a specific task such as the predetermined predictive task.

[0063] Other deep learning models can be used such as, for example: a model based on recurrent neural networks (called "Recurrent Neural Networks" or RNNs), a model based on convolutional neural networks (called "Convolutional Neural Networks" or CNNs), a model based on graph neural networks (called "Graph Neural Network" or GNNs) or even a hybrid model using a combination of transformers, RNNs, CNNs and / or GNNs.

[0064] Step S200 - receipt of the specific labeled dataset:

[0065] Data labeling consists of associating a label with each data point. "The specific dataset is labeled according to the predetermined predictive task" means that the label associated with each data point carries information related to the predetermined predictive task. For example, if the predetermined predictive task is to predict the response to methotrexate in rheumatoid arthritis, according to EULAR criteria, it is possible to determine, for each subject, whether they achieved the EULAR response 6 months after the start of treatment and thus validate the prediction with a dedicated label.

[0066] This data labeling can be carried out manually or automatically either because it is contained in the subject's medical record, or through a feed-forward neural network or a machine learning model such as the Random Forest or the XGboost, for example.

[0067] In the case of applying the present method to research on immune-mediated inflammatory diseases or their management, it is advantageous for the specific dataset to include the medical data of subjects diagnosed or suspected of having an immune-mediated inflammatory disease, and where at least one pathology associated with the predetermined predictive task is an immune-mediated inflammatory disease. This specific dataset has been labeled with respect to the diagnosis of immune-mediated inflammatory diseases and / or their management.

[0068] Step S300 - obtaining the model specific to said at least one pathology:

[0069] Obtaining the model specific to said at least one pathology using an ap Transfer learning can be done in two different ways. The specialization process includes, at your option: - training the pre-trained foundation model on the specific dataset; or - the adaptation of said pre-trained foundation model by the addition of a adapter to the pre-trained foundation model, said adapter being made up of a first neural network trained on said specific dataset.

[0070] In the first case, the specific model is obtained by transferring 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 labeling the specific dataset. This operation is called "fine-tuning" in English.

[0071] Transfer learning is a machine learning technique that allows the transfer of knowledge and skills from a previously trained model (i.e., a pre-trained foundation model) to a new, similar task, thus facilitating learning on new datasets. In this case, the specialization relates to at least one pathology affecting at least one of the subjects in the group. Indeed, this specialization is achieved through training the pre-trained foundation model with specific datasets labeled according to the predetermined predictive task, which itself relates to at least one pathology.This approach allows leveraging prior knowledge and representations learned by the initial foundational model, leading to improved performance and a reduction in the time and resources needed to train a new, specific model from scratch.

[0072] In the second case, the use of an adapter eliminates the need to retrain the pre-trained foundation model, whose layers can then be frozen. The first neural network, which constitutes the adapter, corresponds to a kind of overlay of neurons that is added to the pre-trained transformation model. This modifies its operation by adapting it to the predetermined predictive task, thanks to the training of the adapter on said specific dataset. Moreover, this method is highly modular because, for each predetermined predictive task, it is possible to add a corresponding adapter, where each adapter has been previously trained separately on a dataset 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) being studied, and the medical data available for the group of subjects.

[0073] The integration of the adapter can be done in different ways: by adding the first neural network to the output of the pre-trained foundation model, or within the pre-trained foundation model via the addition of new interlayers medians between its different layers. One way to create an adapter is to use the LORA (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.

[0074] In another advantageous embodiment compatible with the two cases mentioned, and in order to improve and, above all, validate the performance of the model specific to at least one pathology with respect to the predetermined predictive task for which it is trained, the method includes a validation process for this model. For example, this validation can 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 method consists of setting aside a portion of the labeled specific data during training to then be used to evaluate the final model obtained, and this is done iteratively. One or more cost functions can also be used and minimized during training. It is also possible to use Monte Carlo or bootstrap cross-validation.

[0075] It should be noted that training the pre-trained foundation model to obtain the model specific to at least one pathology must be performed specifically for each predetermined predictive task. Consequently, this step is performed for each labeled specific dataset, particularly one relating to different pathologies. Alternatively, this training can also be repeated with another labeled specific dataset, which could be labeled according to a new predetermined predictive task relating to the same pathology as that affecting at least one of the subjects in the group (or relating to another pathology), so as to further specialize the specific model on the same pathology (or to specialize it on two different pathologies).

[0076] The model specific to said at least one pathology obtained at the end of step S300 allows, according to the predetermined predictive task, to make a prediction as to the state of health and / or the health outcome of a subject, based on the medical data received and relating to said subject.

[0077] Step S400 - receipt of medical data relating to the group of subjects:

[0078] As stated previously, medical data generally comprises different modalities (or formats), ranging from free text, such as in medical reports, to tables of numerical data, such as laboratory results, images from various medical imaging devices (e.g., X-rays, CT scans, MRIs, ultrasounds), or even omics (e.g., DNA sequences) or histopathological data. It can therefore be advantageous to preprocess this data by separately encoding the medical data of different modalities.

[0079] By way of illustration and according to a particular embodiment shown in [Fig.2], When the generic, unlabeled dataset includes medical data from at least two different predetermined categories, for each predetermined category, the medical data for that category can be encoded separately from the medical data for another category corresponding to step S130. This encoding step is optional and can be applied in the same way to each of the datasets received during the method's steps for grouping subjects into a group, namely: after receiving the specific labeled dataset and after receiving the medical data for the group of subjects. Therefore, if this encoding is applied to one of the datasets used in the method, it must be applied to the other datasets to ensure the data are processed similarly and, most importantly, in compatible formats.

[0080] As illustrated in [Fig. 3], this encoding is performed by a model specific to each predetermined modality considered. Thus, medical data for modality 1 are encoded by encoder 1, medical data for 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 dataset considered. Preferably, these encodings are performed by models that have been previously trained in a self-supervised manner.

[0081] For example: modality 1 may concern medical data in free text format, modality 2 may concern medical data in table format such as laboratory results / analyses, modality 3 may concern medical images in general (i.e. whatever 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 may concern omics data.

[0082] 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 similar. To achieve this, the laboratory results, like data tables, are converted into free text format (i.e., in the same modality) before being encoded with the data from the medical records, themselves in free text format, by a large language model (called a "Large Language Model" in English) which may include a transformer-type architecture whose operation and technical advantages have been described previously.

[0083] Image data, on the other hand, can be encoded by a model having a visual transformer architecture (called "Visual Transformer - ViT"). 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 the (pre)training of the model specific to image encoding (e.g., 224 x 224). Similarly, histopathological images can be divided into sub-images / patches of the same size (e.g., 224 x 224).

[0084] Advantageously, each input vector is enriched with positional 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 several attention layers as described previously, with importance weights (or attention scores) calculated for each pair of vectors based on their relationships. This allows, in particular, for emphasis to be placed on the parts of the image that are important for the task to be performed.

[0085] Omics / genetic data can be encoded via a convolutional neural network.

[0086] It should be noted that one or more modalities may be removed depending on the data available for the different training sessions as well as in the medical data of the group of subjects to be grouped. The so-called "predetermined" modalities then correspond to the selected / retained modalities, and may be: electronic health records, clinical notes, laboratory results, medical images, omics data and / or histopathological data.

[0087] In another advantageous embodiment, the encoded medical data are then merged by a modal fusion model, see step S140 in [Fig. 2] for the method to obtain a pre-trained foundation model. As with encoding, this encoded data merging step is optional and can be applied in the same way to each of the encoded datasets during the steps of the method for grouping subjects into a group. This merging of the encoded data makes it possible to discover or preserve links / relationships between medical data from different modalities. For this purpose, this merging can be performed by a second neural network with an attention mechanism that will average the vector representations of the encoded medical data.

[0088] Step S500 - creation of the vector representation:

[0089] As explained previously, the Transformer-type foundation model generates a vector representation of the received medical data. With the various training sessions performed, as well as the encoding and / or fusion of the data, this vector representation evolves and becomes more refined.

[0090] In an advantageous embodiment, this creation of the vector representation of medical data uses an algorithm derived from the model specific to said at least one pathology, and where the vector representation of the data medical is representative of information from contributions included in the medical data used to group subjects into said at least one subgroup.

[0091] In order to increase the explainability of the method and thus the explainability of grouping subjects into at least one subgroup, the algorithm derived from the model specific to said at least one pathology is an artificial intelligence explainability algorithm. Depending on the input data, the artificial intelligence explainability algorithm used may be: the SHAP algorithm for "Shapley Additive Explanations," which is best suited when the input data is in array format; AtMan, which provides explanations for Transformer-type models; or Grad CAM, which is best suited when the input data is images. It should be noted that a combination of these algorithms is possible when the input data is of different modalities (i.e., type / format).

[0092] These explainability algorithms make it possible to know the level of contributions of the different medical data which made it possible to make the final prediction in the output of the model specific to said at least one pathology.

[0093] Fig. 4 illustrates a vector representation of the medical data relating to the group of subjects to be grouped into subgroups, and obtained as output of the model specific to said at least one pathology.

[0094] Step S600 - grouping of subjects from the group into subgroups):

[0095] From the vector representation of medical data, it is easy to Group the subjects into at least one subgroup using, for example, distance measurements between vectors. Indeed, thanks to this method, subjects whose predictions are explained in the same way have similar vector representations.

[0096] In the non-limiting example of [Fig.4], it is possible to identify 3 possible subgroups (called "clusters" in English): Cl, C2 and C3.

[0097] As illustrated in [Fig.1] at step S700 and according to an advantageous embodiment, the method includes, after this grouping step S600, the identification of at least one piece of information representative of the explainability of the grouping by the creation of a model for predicting membership in a subgroup allowing to predict the membership, for each of the subjects of the group, of said at least one subgroup.

[0098] Since the links between the medical data of the subjects in the group received at the input and the subgroups formed can be complex, one way to simplify and more easily find information representative of the explainability of the grouping is to use at least one of the following: a statistical test, a rule model such as "skope rules" (acronym for "Supervised Ky Fan Ordering Permutation for Rule Ex") traction” or rule extraction by supervised Ky Fan ordered permutation), or a decision tree model whose branch level can be on the order of 2, as a membership prediction model and which would be previously trained in an unsupervised manner.

[0099] Preferably, said at least one piece of information representative of the explainability of the grouping is at least one biomarker associated with said subgroup of subjects.

[0100] The term “biomarker” here refers to at least one characteristic that can 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. These include, but are not limited to, molecular, histological, radiographic, and physiological biomarkers. The BEST glossary (for “Biomarkers, Endpoints and other Tools”) defines seven categories of biomarkers: susceptibility / risk, diagnostic, monitoring, prognostic, predictive, pharmacodynamic / response, and safety.Finally, the complete identification of a biomarker includes the biomarker name, the source / matrix, the measurable characteristic(s), and the analytical method used to measure the biomarker, which may be a single characteristic or a panel of multiple characteristics.

[0101] In the case of the application of the present method for research on immune-mediated inflammatory diseases, the identification of biomarkers will improve diagnosis, prognosis, and also management by applying treatment adapted to the affected subject.

[0102] The method for grouping a group of subjects into at least one subgroup, as previously described in the different embodiments, can be implemented in a device 200 as illustrated in [Fig.5].

[0103] According to one embodiment, the device 200 comprises at least one processor 202 configured to perform the various steps of the method described above, such as: - step S100: receive a pre-trained foundation model, said pre-trained foundation model having been previously obtained by training, through self-supervised learning, a Transformer type foundation model, on a generic unlabeled dataset; - step S200: receive a labeled specific dataset comprising the medical data of a plurality of subjects, where said labeled specific dataset is labeled according to said predetermined predictive task; - step S300: obtain a model specific to at least one pathology by training said pre-trained foundation model on said set of specific labeled data using transfer learning; - step S400: receive medical data relating to the group of subjects to be grouped into at least one subgroup; - Step S500: Create a vector representation of the medical data relating to the group of subjects received; and - step S600: group subjects from the group into at least one subgroup of subjects according to said vector representation, so as to obtain, for each subject in the group, information on a health status and / or a health outcome, relating to said at least one pathology, according to the grouping of each subject in relation to said at least one subgroup of subjects.

[0104] Advantageously, the processor also includes a generation module, configured to implement the method for obtaining a pre-trained foundation model as described above.

[0105] In a particular embodiment, the processor includes a data processing module configured to encode and / or merge medical data when such data comprises 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 retain and process (or conversely, those that one wishes to delete and not process).

[0106] In another advantageous embodiment, the processor includes an explainability module configured to identify at least one piece of information representative of the explainability of the grouping performed. To this end, the explainability module is capable of creating a subgroup membership prediction model that allows each subject in the group to predict the membership of at least one subgroup, said membership prediction model using at least one of the following: a statistical test, a rule model, or a decision tree model, previously trained in an unsupervised manner.

[0107] According to another embodiment, the device 200 also includes a memory 201 for storing program instructions that can be loaded into the processor 202, such as the execution of the various 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.

[0108] The 202 processor can be, for example: - a processor or processing unit capable of interpreting instructions in a computer language, the processor or processing unit being able to understand, be associated with, or be connected to a memory comprising the instructions instructions, or - the combination of a processor / processing unit and a memory, the processor or processing unit being capable of interpreting instructions in a computer language, the memory containing said instructions, or - an electronic circuit board in which the steps of the invention are described within silicon, or - a programmable electronic chip such as an FPGA chip (for "Field-Programmable Gate Array").

[0109] The device 200 may also include an input interface 203 for receiving input data and an output interface 204 for providing output data. Examples of input data may be medical data relating to the subjects in the group or the various training datasets. Examples of output data may be the subgrouping of each subject, explainability data for the formation of these subgroups such as representative information on the explainability of the grouping, or the vector representation of the medical data.

[0110] To facilitate interaction with the device 200, a screen 205 and a keyboard 206 may be provided and connected to the computer process / circuit 202.

[0111] This disclosure also relates to a computer program product intended to provide 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 execute the steps of the method according to one of the embodiments described above.

[0112] The computer program product for executing the method described above may be written as computer programs, code segments, instructions, or any combination thereof, to instruct or configure the processor or computer individually or collectively to function as a special-purpose machine or computer to perform the operations carried out 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 with ordinary skills in the art can easily write instructions or software based on the functional diagrams and flowcharts illustrated in the drawings and corresponding descriptions in the specification, which disclose algorithms to perform the operations of the method as described above.

[0113] This 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.

[0114] According to one embodiment, the computer-readable storage medium is a non-transient computer-readable storage medium.

[0115] Computer programs implementing the method described in these embodiments can generally be distributed to users on a computer-readable storage medium, such as an SD card, an external storage device, an electronic chip, a flash memory device, a portable hard drive, and software websites, but not limited to these. From the distribution medium, the computer programs can be copied to a hard drive or similar intermediate storage medium. The computer programs can be executed by loading computer instructions, either from their distribution medium or from their intermediate storage medium, into the computer's execution memory, configuring the computer to act in accordance with the method of this disclosure.All these operations are well known to people with expertise in computer systems.

[0116] The instructions or software intended to control a processor or computer to implement the hardware components and execute the methods described above, together with all associated data, data files and data structures, are recorded, stored or fixed in or on one or more non-transient storage media readable by a computer. Examples of non-transient storage media readable by a computer 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-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 drives,Semiconductor disks, and any device known to a person skilled in the art capable of storing instructions or software and all associated data, data files, and data structures in a nontransient 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 networked coupled computer systems, so that the instructions, 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

1. Claims A device (200) 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 (202) configured to: 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; b. receiving a labeled specific data set comprising medical data of a plurality of subjects, wherein said labeled specific data set is labeled according to said predetermined predictive task (S200); c. obtaining a model specific to said at least one pathology by a specialization process via transfer learning (S300), said specialization process consisting of: • training said pre-trained foundation model on said specific data set; 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. receive medical data relating to the group of subjects to be grouped into at least one subgroup (S400); e. create a vector representation of the medical data relating to the group of subjects received (S500); and f. 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 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 (S600).

2. Device (200) according to claim 1, wherein, when the medical data 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 a model specific to the predetermined modality considered, said model specific to the predetermined modality considered having been previously trained in a self-supervised manner (S20).

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 (S30).

4. The device (200) of 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 creation of said vector representation of the medical data (S500) uses an algorithm derived from the model specific to said at least one pathology, and where the vector representation of the medical data is representative of contribution information included in the medical data used to group the subjects into said at least one subgroup.

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

7. Device (200) according to any one of the preceding claims, comprising, after grouping the subjects into at least one subgroup (S600), identifying at least one piece of information representative of the explainability 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 (S700).

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

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 subject and / or prediction of response to a treatment.

11. Device (200) according to one of the preceding claims, also comprising a process for validating the model specific to said at least one pathology.

12. Device (200) according to one of the preceding claims, wherein the specific data set comprises the medical data of subjects 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 subject.

Images