Method for predicting a mental state of a subject based on its individual dynamics and system

Abstract

A method for generating with at least one machine learning algorithm (ML) a prediction of at least one indicator of a mental state of a living subject likely to suffer from a mental disorder, the method comprising: a. acquiring at least one set of time-dependent physiological features of the subject, possibly together with at least one static variable of the subject, b. determining for at least one predetermined hyperparameter ν corresponding to a number of input features for the at least one machine learning algorithm (ML), a specific value ν1 of ν for the subject, c. selecting ν1 input features for the at least one ML algorithm amongst the acquired physiological features, the ν1 input features forming a biosignature of the subject, d. generating, using the at least one machine learning (ML) algorithm, a prediction of the at least one indicator of the mental state of the subject.

Description

[0001] Description

[0002] Title: Method for predicting a mental state of a subject based on its individual dynamics and system

[0003] Field of the invention

[0004] The present disclosure generally relates to methods and systems for generating predictions of a mental state of a living subject based on time-dependent data.

[0005] Background of the invention

[0006] Machine Learning (ML) has recently been demonstrated to rival expert-level human accuracy in prediction and detection tasks in a variety of domains, including healthcare. As the need for automated health care solutions is increasingly gaining interest, the vision of being able to accurately predict a patient's medical trajectory by integrating vast amounts of disparate data has inspired generations of computer scientists. It has resulted in a variety of early applications of artificial intelligence (Al) in the form of decision aids and decision support systems.

[0007] Major depressive disorder (MDD) is a prevalent, disabling and chronic, biologically- based disorder that impairs social and educational functioning. Individuals with MDD experience a heterogeneous range of symptoms, which can change over time. The recovery time of major depressive episodes is also highly variable at the individual level. A diagnosis of MDD currently relies solely on subjective markers such as statements made by a clinician based on the patient narrative, which presents numerous challenges in predicting treatment response, remission, risk of relapse or recovery.

[0008] Many studies have therefore investigated more objective markers like physiological measurements including heart rate variabilities, psychomotor retardation, or sleep alteration.

[0009] Rykov et al., in the article “Digital Biomarkers for Depression Screening With Wearable Devices: Cross-sectional Study With Machine Learning Modeling” , JMIR Mhealth Uhealth 9 (2021), developed a signature to predict the risk of depression, based on physiological and behavioral data characterizing physical activity, sleep patterns, and circadian rhythms recorded by a wearable device. Despite great potential, the signature unfortunately showed limited predictive ability.

[0010] Following a different approach, Lewis et al. used, in the abstract "Mixed Effects Random Forests for Personalised Predictions of Clinical Depression Severity", ICML 2021, mixed effects random forest models trained on physiological and behavioral data to predict depression severity. They reported improved performance compared to the literature that they attributed to the ability of mixed effects random forest models to personalize some of the model parameters, accounting for the random effect in the model, to individuals tested.

[0011] However, given the heterogeneous nature of patients’ symptoms and variable recovery time, it is difficult to achieve high levels of prediction accuracy using models with limited personalization, as showed in mixed effects random forest models, and based on the same set of features for all potential patients.

[0012] There is therefore a need for an improved prediction method based on a patient- specific signature that would take as input specific features for each patient, characterizing their individual symptoms.

[0013] Summary

[0014] The present invention aims to remedy to all or part of the deficiencies of the prior art mentioned above and exemplary embodiments of the invention relate to a prediction method for generating with at least one machine learning algorithm (ML) a prediction of at least one indicator of a mental state of a living subject likely to suffer from a mental disorder, the method comprising: a. acquiring at least one set of time-dependent physiological features of the subject, possibly together with at least one static variable of the subject, b. determining for at least one predetermined hyperparameter v corresponding to a number of input features for the at least one machine learning algorithm (ML), a specific value vl of v for the subject, c. selecting vi input features for the at least one ML algorithm amongst the acquired time-dependent physiological features, the Vi input features forming a bio signature of the subject, d. generating, using the at least one machine learning (ML) algorithm, a prediction of the at least one indicator of the mental state of the subject.

[0015] The determination for at least one predetermined hyperparameter v of a specific value vi of v for the subject in step b) may comprise the use of statistical methods and / or machine learning techniques. In some non-limiting embodiments, it involves an optimization process based on the predictions generated by the at least one ML algorithm of step d) configured by said specific value or values to determine, for a group of other living subjects than the subject, as will be detailed hereafter. The selection in step c) of Vi input features for the at least one ML algorithm amongst the acquired time-dependent physiological features may comprise the use of statistical methods and / or machine learning techniques. Preferentially it involves selecting the Vi features most correlated with training indicator labels of the subject, as will be detailed hereafter.

[0016] Living subject

[0017] A “living subject” refers to any living one for whom the system and methods of the invention may be implemented. Preferentially, the living subject is a human being, though the present disclosure is not limited to humans, but may also include other types of mammals or other animals, such as pets.

[0018] By “likely to suffer from a mental disorder”, it is meant that the living subject, referred to as a “patient” in some embodiments, has not necessarily been diagnosed as suffering from a mental disorder.

[0019] Mental disorder

[0020] The invention is not limited to a particular type of mental disorder.

[0021] Preferentially, the mental disorder comprises a mood disorder and / or an anxiety disorder. In particular, the mental disorder may comprise major depressive disorder (MDD).

[0022] Indicator

[0023] By “indicator of a mental state of a subject”, it is meant that the indicator alone may provide enough information for the decision-maker and / or the user of the system of the invention to evaluate the mental state of the subject. Preferentially, the indicator of the mental state of the subject is based on a standardized clinical scale known from the art for the mental disorder considered, on which health recommendations may be established, such as the Montgomery-Asberg Depression Rating Scale (MADRS) for MDD.

[0024] Time-dependent physiological features

[0025] By “time-dependent physiological features”, it is meant a collection of physiological measurements obtained through at least two repeated measurements of the subject during a monitoring period.

[0026] The present invention is not limited to a particular monitoring period, or to a particular time period over which measurements are acquired. The monitoring period may range from a few weeks to a few months. The duration may be shorter or longer if possible or needed. The time-dependent features may be collected at different sample rates, depending for example on the acquisition device used for the measurements.

[0027] The time-dependent physiological features acquired can include features expressing physical activity and / or heart rate and / or heart rate variability and / or breathing rate and / or sleep, and preferentially all of them.

[0028] Static variables

[0029] By “static variable”, it is meant variables that do not change or change very slightly through the monitoring period during which the time-dependent physiological features are acquired.

[0030] The static variables may be specific to each subject and may comprise its age, gender, treatment regimen or socioeconomic status.

[0031] The static variables are only used during the determination step b) of the prediction method. They are not used to generate the prediction, using the at least one ML algorithm, of the at least one indicator of the mental state of the subject.

[0032] Input features

[0033] “Input features” refer to the features amongst the set of acquired time-dependent physiological features that are used by the at least one ML algorithm to generate a prediction. All or part of the acquired physiological features can be used, their number being set by the number of input features Vi.

[0034] In some non-limiting embodiments, the Vi features selected are the Vi features most correlated, using preferentially the Hilbert- Schmidt Independence Criterion (HSIC), with training indicator labels acquired on the subject and used to train the at least one ML algorithm.

[0035] The Vi selected features which are specific to the subject form a biosignature of the subject.

[0036] Predetermined hyperparameters

[0037] The term “hyperparameter” in the present invention can refer to model hyperparameters that affect the model selection task of the at least one ML algorithm, such as the topology, the size of a neural network, the number of input features or the way training labels are processed during the training of the model.

[0038] By “predetermined”, it is meant that the at least one hyperparameter must be fixed before implementing the method of the present disclosure. The choice of hyperparameters constitutes an embodiment of the present invention. The determination step b) only chooses the specific values for the subject of a given set of hyperparameters.

[0039] Several hyperparameters can be chosen but the number of input features v must always be included amongst them. When only one hyperparameter is fixed, it is the number of input features v.

[0040] In some non-limiting embodiments, the set of predetermined hyperparameters comprises a constant of an optimistic model X, called the healing rate, for which a specific value i is determined for the subject, the at least one ML algorithm being trained using residual indicator labels of the subject, generated from the training indicator labels acquired on the subject by the optimistic model configured by i.

[0041] Preferentially, there are two predetermined hyperparameters, v the number of input features for the at least one ML algorithm, and X.

[0042] Training indicator labels

[0043] Training indicator labels refer to evaluations of the indicator of the mental state of the subject at different time points during the monitoring period, that are used to train the at least one ML algorithm.

[0044] As mentioned above, in some non-limiting embodiments, the set of predetermined hyperparameters comprises the healing rate of an optimistic model X, for which a specific value i is determined for the subject, the at least one ML algorithm being trained using residual indicator labels of the subject, generated from the training indicator labels by the optimistic model configured by i.

[0045] By “residual indicator”, it is meant the difference between the training indicator and the result of the “optimistic model”. The optimistic model predicts the evolution of an indicator based solely on i and a previous evaluation of the indicator.

[0046] The optimistic model can follow the equation yd where ydis the indicator label of time point dtand ydis the indicator label of time point d. with dL< d. In particular, dtand d are days.

[0047] ML algorithm

[0048] In some non-limiting embodiments, the at least one ML algorithm is configured using the number of input features Vi and / or another specific value for the subject of at least one other predetermined hyperparameter. By “configured”, it is meant arranged in its functional units, in its chief characteristics, in the way it processes input data.

[0049] In particular, in some non-limiting embodiments, the at least one ML algorithm is a multilayer perceptron (MLP) whose depth and / or number of neuron units per layer is a function of the specific value vi for the subject and / or is a function of said specific value for the subject of said at least one other predetermined hyperparameter.

[0050] Determination of the hyperparameter values

[0051] The determination for at least one predetermined hyperparameter v of a specific value vi of v for the subject may comprise the use of statistical methods and / or machine learning techniques.

[0052] In some non-limiting embodiments, the determination of the specific value for the subject of the at least one predetermined hyperparameter v, preferentially the determination of the specific values vi and i, involves at least one optimization process based on time-dependent physiological features, possibly together with at least one static variable, collected from a group of other living subjects or a group of other living subjects and the subject. In the non-limiting embodiments where there are several predetermined hyperparameters, the determination of the specific values for the subject may involve separate optimization processes for every predetermined hyperparameter.

[0053] In some non-limiting embodiments, the determination of the specific value for the subject of the predetermined hyperparameter or hyperparameters is carried out by an optimization process maximizing a metric measuring the accuracy and / or minimizing a metric measuring the error of a prediction, the prediction being generated through the at least one ML algorithm configured by the said specific value or said specific values to determine, and trained and evaluated using data collected from a group of other living subjects than the subject.

[0054] Preferentially, during the optimization process, the ML algorithm can be trained and evaluated using a Leave-one-Patient-Out approach.

[0055] A grid search can be used during the optimization process to select potential hyperparameter values.

[0056] Training method

[0057] Another aspect of the invention relates to a method for training the at least one ML algorithm of the prediction method as defined above, to generate a prediction of the at least one indicator of the mental state of a living subject p likely to suffer from a mental disorder, the training method comprising:

[0058] • generating a training dataset using data obtained from the subject p and a group of other living subjects P, said generating comprising for each subject: o acquiring during a monitoring period, time-dependent physiological features from at least one acquisition device sensing the subject, possibly together with at least one static variable of the subject, o generating training indicator labels, said labels expressing the mental state of said subject at different time points during the monitoring period,

[0059] • determining for the subject p for a at least one predetermined hyperparameter v corresponding to a number of input features for the at least one ML algorithm, a specific value vi of v, based on the data collected from the group of other living subjects P or the group of other living subjects P and the subject p,

[0060] • using only the training data of the subject p: o selecting vi input features amongst the physiological features acquired, o training the at least one ML algorithm using the training indicator labels and input features.

[0061] A further aspect of the invention relates to a computer-based system implementing the prediction method as defined above, comprising a processing unit for determining for at least one predetermined hyperparameter v corresponding to a number of input features for at least one machine learning algorithm (ML) a specific value vi of v for the subject, training the at least one ML algorithm according to the training method as defined above and generating a prediction of the at least one indicator of the mental state of the subject through the at least one ML algorithm.

[0062] Computer-based systems

[0063] By “computer-based”, it is meant that the above-described system and method according to the invention may be implemented using any computer system.

[0064] The computer system may include any automated system in the form of any hardware and / or software, such as a laptop, a personal computer, a smartphone, a workstation, a computer terminal, one or several network computers, or any other data processing system or user device. Preferentially, the computer system comprises at least one processing unit which controls the overall operation related to the method of the invention by executing computer program instructions. The processing unit may comprise general or special purpose microprocessors, microcontrollers or both, or any other kind of CPU, GPU, FPGA, quantum processor and any combination of those.

[0065] The computer system may comprise one or more storage media on which computer program instructions and / or data are stored. The storage media may comprise magnetic, magneto-optical disks, optical disks, flash memory devices, solid state drives (SSD), cloud storage, or any other type of removable storage medium or a combination of both.

[0066] The computer system preferably comprises at least one memory unit that can be used to load the instructions or / and data when they are to be executed by the processing unit.

[0067] The memory unit comprises for instance a random-access memory (RAM) and / or a read only memory (ROM).

[0068] The computer system may further comprise one or more user interfaces, such as any type of display, a keyboard, a pointing device such as a mouse or trackpad, audio input devices such as speakers or microphones, or any other type of device that allows user interaction with the computer system.

[0069] The computer system may comprise one or more network interfaces for communicating with other devices.

[0070] User- interface

[0071] In certain non-limiting embodiments, the system of the invention may further comprise a user-interface for informing a user of the prediction generated of the at least one indicator of the mental state of the subject and / or of the selected input features forming a biosignature of the subject via an image and / or text.

[0072] The term “user” as used herein includes, for example, a person or entity that owns a user computer, such as a computing device or a wireless device; a person or entity that operates or utilizes a user computer; or a person or entity that is otherwise associated with a user computer. It is contemplated that the term “user” is not intended to be limiting and can include various examples beyond those described.

[0073] The user may be a different person from the living subject. The user may be a decisionmaker of various nature. Preferentially, the user is a clinician and the living subject a patient of the clinician. Acquisition device

[0074] The system of the invention may comprise an acquisition device sensing the subject for generating time-dependent features.

[0075] In some embodiments, such an acquisition device comprises a wearable device, such as a wrist monitor, a smart watch, a smart ring, a smart collar, a smart belt, and / or smart patch.

[0076] The acquisition device may further comprise other types of smart devices, such as a smart bed, a smart phone, or any other connected device suitable for sensing the subject.

[0077] Preferentially, the acquisition device comprises at least one of an accelerometer, photoplethysmography (PPG), an optical sensor, a heart rate detector (e.g., through optical detection or electrical detection) and an electrodermal activity sensor (e.g., through a measurement of dermal electrical impedance or dermal capacity), or electromagnetic sensors such as UWB position sensors.

[0078] Brief description of the drawings

[0079] For a more complete understanding of the present invention, a description will now be given of several examples, taken in conjunction with the accompanying drawings, in which:

[0080] [Fig 1] illustrates different steps of an example of a prediction method according to the invention.

[0081] [Fig 2] depicts an example of a temporal evolution of a MADRS score recorded on a subject suffering from MDD.

[0082] [Fig 3] illustrates parts of the training process of a ML algorithm to generate a prediction of an indicator of a mental state of a subject according to the invention.

[0083] [Fig 4] shows different steps of an example of the determination process to choose a specific value for a subject of a set of hyperparameters.

[0084] [Fig 5] reports in a table the specific values selected for 26 different patients, of a set of hyperparameters ( , v) where is the healing rate of an optimistic model and v a number of input features.

[0085] [Fig 6] represents the origin of the input features selected by two different patients to generate predictions.

[0086] Detailed description

[0087] Using the figures 1 to 4, different steps of an example of a prediction method using at least one ML algorithm and an example of a method to train the ML algorithm according to the invention are described. The figure 1 illustrates four steps of an example of a prediction method according to the invention.

[0088] The step 1 includes the acquisition of a set of time dependent physiological features of the subject during a monitoring period and the recording of corresponding indicator labels evaluated at different time points during the monitoring period (figure 2).

[0089] The step 2 comprises the determination for at least one predetermined hyperparameter v corresponding to a number of input features for the at least one ML algorithm, a specific value vi of v for the subject using machine learning and / or statistical techniques (figure 4).

[0090] The step 3 includes the selection of vi input features amongst the physiological features acquired using statistical methods and / or machine learning techniques (figure 3).

[0091] The final step involves generating a prediction using the at least one ML algorithm based on the vi input features selected, of the at least one indicator of the mental state of the subject.

[0092] The figure 2 illustrates an example of the recording of indicator labels of a subject wherein the mental disorder comprises MDD and the indicator is the MADRS score.

[0093] The points 1 in figure 2 represent clinical evaluations of the mental state of the subject made by a clinician during monthly clinical visits through a six-month period. These clinical evaluations constitute indicator labels.

[0094] The line 2 is the real evolution of the MADRS score of the subject that is unknown at the exception of the days of the clinical visits.

[0095] The clinical evaluation made on the day d of the clinical visit each month can be extended in a confidence region around the visit, depending on the type of data labelled and the labels periodicity. In 3, this extension period is chosen to be 5 days before and five days after the clinical visit. This extension methodology can be justified by considering the test-retest reliability of the MADRS score over the course of several days as described in the literature. It helps mainly to generate more training indicator labels for the training of the ML algorithm.

[0096] The line 4 is the output of an optimistic model, which estimates the MADRS evolution based solely on the most recent clinical visit and a constant X, called the healing rate. On a given day, this model predicts a MADRS value given an amelioration rate based on the previous clinical evaluation of the patient by the clinician. The difference between the actual MADRS and the MADRS value predicted by this optimistic model is called the residual MADRS, represented in 5. The six-month period can be separated into a training period 6, wherein the indicator labels acquired are used to train the ML algorithm, and an evaluation period 7, wherein the indicator labels are used to evaluate the ML algorithm.

[0097] The figure 3 illustrates the third and final part of the training process of the at least one ML algorithm to generate a prediction of an indicator of a mental state of a subject p, in an exemplary embodiment of the invention where the predetermined hyperparameters are a couple (X, v) where Z. is the healing rate of an optimistic model and v a number of input features for the at least one ML algorithm.

[0098] During a first part of the training process, illustrated in figure 2, a dataset was collected on the subject p during a six-month period: a set of physiological features was acquired daily during the period, an indicator label was evaluated at a first clinical visit at day d0and at monthly visits dlrd2, d3, d4, d5, d6during the six months. The evaluations were extended to 5 days before and after the clinical visits.

[0099] During a second part of the training process, described hereafter in figure 4, the specific values ( i, Vi) for the subject p of the couple (X, v) were determined.

[0100] At the beginning of the third and final step, the dataset acquired on the subject p is preprocessed. The data from the first three months are used as training / validation dataset and the remaining data as test set:

[0101] XTrain / vai’ ^Train / val ^d<d4> ^d<d4^

[0102] ^Test’ ^Test ~ d>d4> d>d4where every data points (x, y) consists of the physiological measures at a given day d and its associated MADRS value y. The training / validation dataset is further decomposed into a training set (first 80%) and validation set (last 20%). The training set is used to parameterize the model, validation is used to control overfitting, and the test set is only used to compute the metrics and evaluate the model’s performance.

[0103] At a selection step S, the Vi features amongst the set of features XTrain / vaimost correlated with the labels YTrain / vaiusing the Hilbert- Schmidt Independence Criterion (HSIC) are selected, forming a new input matrix XTrain / vai. The indicator labels YTrainare fed into an optimistic model O, configured by i of the form yd where ydis the indicator label of day d; and ydis the indicator label of day d, with dt< d. The optimist model O outputs residual indicator labels R and base indicator labels B.

[0104] Using the selected features XTrain / vaiand the residual indicator labels R, a ML algorithm, which is a multi-layer perceptron (MLP) is trained to output an estimate P of the residual indicators R. Specifically, the MLP consists of an input of dimension vi followed by 3 hidden layers of respectively 8vi, 4vi and 2vi neurons, and an output of dimension 1 which is the scalar prediction.

[0105] The parameters chosen for MLP training are : batch size of 16, training for 500 epochs, and early stopping callback of 5 epochs monitoring improvements in validation loss. This model is trained on the train dataset and validated on the validation dataset as defined previously. To smooth out random fluctuations due to kernel initialization, and to avoid having inaccurate predictions because of potential local minima in the parameter space of the model, this process is repeated 11 times and the final output prediction is set to be the median of the predictions.

[0106] To give the complete MADRS score prediction I (figure 3), the output of the MLP is summed with the base indicator labels B.

[0107] The optimistic model configured by i, the selection step of vi features and the MLP configured by vi form a specific model, that is called the Signature Based Model of Depression (SiBaMoD) of patient p: SiBaMoDP(ki, vi). It is represented in 12b figure 3.

[0108] The figure 4 illustrates in 10 the step of determining for a subject p using an optimization process, the specific values ( i, vi) of the predetermined hyperparameters ( , v) in an exemplary embodiment of the invention, SiBaMoD^ki, Vi) where is the healing rate of an optimistic model and v a number of input features.

[0109] As illustrated in figure 2, a dataset was collected on a group of other living subjects P than the subject p during a six-month period. For each subject: a set of physiological features was acquired daily during the period, an indicator label was evaluated at a first clinical visit at day d0and at monthly visits dlrd2, d3, d4, d5, d6during the six months. The evaluations were extended to 5 days before and after the clinical visits. Using a grid search 11 (figure 4) of hyperparameter values of (X, v) with 0.2< A < 2.7 and 2< v < 101, a potential set of values is defined (X’, v’) at a step 12a. A SiBaMoD^ (X’, v’) is trained as defined in figure 3 for each subject p ’ of P in a Leave-one-Patient-Out approach at a step 12b and an evaluation of the prediction on all the patients p’ of P is generated using binary accuracy at a step 12c. The values selected are :

[0110] The figure 5 is a table that reports examples of a set of specific values selected for 26 subjects in an exemplary embodiment of the invention, SiBaMoD(X, v) wherein the mental disorder comprises MDD and the indicator is the MADRS score. Most patients have selected the same set of values, but 4 patients have different sets.

[0111] The figure 6 illustrates the origin of the features selected by two different patients to form their specific signatures. The circular histogram represents the importance of each group of features in the patient’s signature, weighted by the number of features by categories. For patient 1, features describing disturbances in sleep are the most represented whereas for patient 5, these are the features describing the breathing rate. As can be seen from the examples, the biosignature captures the heterogeneity of the expression of the depression at the individual level, and therefore the potential of the biosignature to orientate treatment in a more personalized manner.

[0112] The invention is not limited to the examples that have just been described.

[0113] Other types and number of predetermined hyperparameters may be chosen. The determination of specific values for a subject may rely on statistical techniques instead of using the ML algorithm of the prediction step. Static variables can be collected on the subjects alongside the physiological features and used during the determination step.

Claims

Claims1. A method for generating with at least one machine learning algorithm (ML) a prediction of at least one indicator of a mental state of a living subject likely to suffer from a mental disorder, the method comprising: a. Acquiring at least one set of time-dependent physiological features of the subject, possibly together with at least one static variable of the subject, b. determining for at least one predetermined hyperparameter v corresponding to a number of input features for the at least one machine learning algorithm (ML), a specific value vi of v for the subject, c. selecting vi input features for the at least one ML algorithm amongst the acquired time-dependent physiological features, the vi input features forming a biosignature of the subject, d. generating, using the at least one machine learning (ML) algorithm, a prediction of the at least one indicator of the mental state of the subject.

2. The method according to the previous claim, the vi features selected being the vi features most correlated, using preferentially the Hilbert- Schmidt Independence Criterion (HSIC), with training indicator labels acquired on the subject and used to train the at least one ML algorithm.

3. The method according to any of the previous claims, the predetermined hyperparameters including a constant of an optimistic model A, called the healing rate, for which a specific value i is determined for the subject, the at least one ML algorithm being trained using residual indicator labels of the subject, generated from the training indicator labels acquired on the subject by the optimistic model configured by i.

4. The method according to the previous claim, the predetermined hyperparameters being X and v.

5. The method according to any of the two previous claims, the optimistic model being of the form yd=where ydis the indicator label of day dLand ydis the indicator label of day d, with dt< d.

6. The method according to any of the previous claims, the at least one ML algorithm being configured using the number of input features vi and / or another specific value for the subject of at least one other predetermined hyperparameter.

7. The method according to the previous claim, the at least one ML algorithm being a multilayer perceptron (MLP) whose depth and / or number of neuron units per layer is afunction of the specific value Vi for the subject and / or is a function of said specific value for the subject of said at least one other predetermined hyperparameter.

8. The method according to any of the previous claims, wherein the determination of the specific value for the subject of the predetermined hyperparameter or hyperparameters, is carried out by an optimization process maximizing a metric measuring the accuracy and / or minimizing a metric measuring the error of a prediction, the prediction being generated through the at least one ML algorithm configured by the said specific value or said specific values to determine, and trained and evaluated using data collected from a group of other living subjects than the subject.

9. The method according to the previous claim, wherein the at least one ML algorithm is trained and evaluated using data collected from the group of other living subjects with a Leave-one-Patient-Out approach.

10. The method according to any of the two previous claims, a grid search being used during the optimization process to select potential hyperparameter values.

11. The method according to any of the previous claims, the mental disorder being major depressive disorder (MDD) and the at least one indicator being the Montgomery-Asberg Depression Rating Scale (MADRS).

12. The method according to any of the previous claims, the physiological features including features expressing physical activity and / or heart rate and / or heart rate variability and / or breathing rate and / or sleep, preferentially all of them.

13. A method for training the at least one ML algorithm of the method according to any of the previous claims, to generate a prediction of at least one indicator of a mental state of a living subject p likely to suffer from a mental disorder, the training method comprising:• generating a training dataset using data obtained from the subject p and a group of other living subjects P, said generating comprising for each subject: o acquiring during a monitoring period, time-dependent physiological features from at least one acquisition device sensing the subject, possibly together with at least one static variable of the subject, o generating training indicator labels, said labels expressing the mental state of said subject at different time points during the monitoring period,• determining for the subject p for a at least one predetermined hyperparameter v corresponding to a number of input features for the ML algorithm, a specific valuevi of v, based on the data collected from the group of other living subjects P or the group of other living subjects P and the subject p,• using only the training data of the subject p~. o selecting vi input features amongst the physiological features acquired, o training the at least one ML algorithm using the training indicator labels and input features.

14. A computer-based system for implementing the method according to any of the claims 1 to 12, comprising a processing unit running a program for determining for at least one predetermined hyperparameter v corresponding to a number of input features for the at least one ML a specific value vi of v for the subject, training the at least one ML algorithm according to the claim 13 and generating a prediction of the at least one indicator of the mental state of the subject through the at least one ML algorithm.

15. The system according to the previous claim, the system comprising a user interface for informing a user, preferentially a clinician, of the prediction generated and / or the input features selected to generate the prediction via an image and / or text.

16. The system according to any of the two previous claims, the system comprising an acquisition device sensing the subject during a monitoring period to acquire a set of time-dependent physiological features.

17. The system according to the previous claim, the acquisition device comprising at least one of a photoplethysmography (PPG), an accelerometer and an electrodermal activity sensor.

Images