Method for predicting mental states based on a subject's individual dynamics and systems

Abstract

1. A method of generating a prediction of at least one indicator of a mental state of a living subject at risk of suffering from a mental disorder using at least one machine learning algorithm (ML), comprising: a. acquiring at least one set of time-dependent physiological characteristics of the subject, optionally together with at least one static variable of the subject; b. for at least one predetermined hyperparameter v corresponding to a number of input features for said at least one machine learning (ML) algorithm, determining a particular value v of v for said subject; c. selecting v1 input features for the at least one ML algorithm from the acquired time-dependent physiological features, wherein the v1 input features form a biosignature of the subject; d. using the at least one machine learning (ML) algorithm to generate a prediction of the at least one indicator of the mental state of the subject. The method comprising:

Description

[Technical Field]

[0001] The present disclosure generally relates to methods and systems for generating predictions of mental states of living subjects based on time-dependent data. [Background technology]

[0002] Machine learning (ML) has recently demonstrated the ability to match expert-level human accuracy in prediction and detection tasks across a variety of domains, including healthcare. With the ever-increasing need for automated healthcare solutions, the vision of being able to accurately predict a patient's medical progress by integrating vast amounts of disparate data has inspired generations of computer scientists. This has resulted in a variety of early applications of artificial intelligence (AI) in the form of decision support and decision assistance systems.

[0003] Major depressive disorder (MDD) is a biologically-based disorder associated with widespread, chronic impairment that impairs social and educational functioning. People with MDD experience a variety of symptoms that may change over time. The duration of recovery from a major depressive episode also varies widely at the individual level. Diagnosis of MDD currently relies solely on subjective indicators, such as statements made by clinicians based on patient narratives, which presents many challenges in predicting treatment response, remission, risk of relapse, or recovery.

[0004] Therefore, many studies have investigated more objective indicators such as physiological measures including heart rate variabilities, psychomotor retardation, or sleep alteration.

[0005] In their paper "Digital Biomarkers for Depression Screening With Wearable Devices: Cross-sectional Study With Machine Learning Modeling," JMIR Mhealth Uhealth 9 (2021), Rykov et al. developed a signature for predicting depression risk based on physiological and behavioral data characterizing physical activity, sleep patterns, and circadian rhythms recorded by wearable devices. Despite the great potential of this signature, unfortunately, it showed limited predictive ability.

[0006] Lewis et al. took a different approach and, in their paper "Mixed Effects Random Forests for Personalized Predictions of Clinical Depression Severity," ICML 2021 (abstract), used mixed effects random forest models trained on physiological and behavioral data to predict depression severity. They reported that mixed effects random forest models performed better than existing literature because they account for random effects within the model, contributing to personalizing some of the model parameters to suit the people being tested. Summary of the Invention [Problem to be solved by the invention]

[0007] However, given the heterogeneous nature of patient symptoms and variability in recovery times, it is difficult to achieve a high level of predictive accuracy when using models with limited personalization capabilities and based on the same feature set for all potential patients, as seen in mixed-effects random forest models.

[0008] Therefore, there is a need for improved prediction methods based on patient-specific signatures that take as input specific features for each patient and characterize each patient's individual condition. [Means for solving the problem]

[0009] The present invention aims to solve all or some of the problems of the prior art discussed above, and an exemplary embodiment of the present invention is a method of generating a prediction of at least one indicator of a mental state of a living subject likely to suffer from a mental disorder using at least one machine learning algorithm (ML), the method comprising: a. acquiring at least one set of time-dependent physiological features of the subject, optionally together with at least one static variable of the subject; b. For at least one predetermined hyperparameter v corresponding to a number of input features for the at least one machine learning (ML) algorithm, determining a particular value v1 of v for the subject; c. selecting v input features for the at least one ML algorithm from the acquired time-dependent physiological features, wherein the v input features form a biosignature of the subject; d. using the at least one machine learning (ML) algorithm to generate a prediction of the at least one indicator of the mental state of the subject. The present invention relates to the above method, which comprises:

[0010] For at least one predetermined hyperparameter v, step b) determining a specific value v1 of v for the subject may include the use of statistical methods and / or machine learning techniques, which in some non-limiting embodiments include an optimization process based on predictions generated by at least one ML algorithm of step d) configured to determine one or more of the specific values ​​for a group of other living subjects other than the subject, as described in more detail below.

[0011] Step c) of selecting v1 input features for the at least one ML algorithm from the obtained time-dependent physiological features may comprise using statistical methods and / or machine learning techniques, preferably comprising selecting v1 features that are most correlated with the training indicator label of the subject.

[0012] Living subjects

[0013] "Living subject" refers to any living subject in which the systems and methods of the present invention may be implemented. Preferably, the living subject is a human, although the present disclosure is not limited to humans and may also include other types of mammals or other animals, such as pets.

[0014] By "potentially suffering from a mental disorder" it is meant, in some embodiments, that a living subject referred to as a "patient" has not necessarily been diagnosed as suffering from a mental disorder.

[0015] Mental disorders

[0016] The present invention is not limited to any particular type of mental disorder.

[0017] Preferably, the psychiatric disorder comprises a mood disorder and / or an anxiety disorder. In particular, the psychiatric disorder may comprise major depressive disorder (MDD).

[0018] index

[0019] By "an index of a subject's mental state," it is meant that the index alone may provide sufficient information to a decision-maker and / or user of the system of the present invention to assess the subject's mental state. Preferably, the index of a subject's mental state is based on a standardized clinical measure known in the art for the psychiatric disorder under consideration, e.g., the Montgomery-Asberg Depression Rating Scale (MADRS) for MDD, on the basis of which health recommendations may be established.

[0020] Time-dependent physiological characteristics

[0021] By "time-dependent physiological characteristic" is meant a set of physiological measurements obtained through at least two repeated measurements of the subject during a monitoring period.

[0022] The present invention is not limited to any particular monitoring period or period over which measurements are taken. The monitoring period may range from weeks to months. The period may be shorter or longer if possible or necessary.

[0023] The time-dependent features may be collected at different sample rates, for example depending on the acquisition device used for the measurements.

[0024] The acquired time-dependent physiological features may include features representative of physical activity and / or heart rate and / or heart rate variability and / or respiratory rate and / or sleep, preferably all of them.

[0025] Static Variables

[0026] By "static variable" is meant a variable that does not change or changes very little throughout the monitoring period during which time-dependent physiological characteristics are acquired.

[0027] The static variables may be specific to each subject and may include age, sex, treatment regimen, or socioeconomic status.

[0028] The static variables are only used during determination step b) of the prediction method: they are not used by the at least one ML algorithm to generate a prediction of the at least one indicator of the subject's mental state.

[0029] Input Features

[0030] "Input features" refer to features from a set of acquired time-dependent physiological features that are used by the at least one ML algorithm to generate predictions. All or some of the acquired physiological features may be used, the number of which is set by the number of input features v1.

[0031] In some non-limiting embodiments, the v1 features selected are the v1 features that are most correlated with the training index labels obtained for the subject and used to train the at least one ML algorithm, preferentially using the Hilbert-Schmidt Independence Criterion (HSIC).

[0032] The selected features v1 that are unique to the subject form the subject's biosignature.

[0033] Given hyperparameters

[0034] The term "hyperparameter" in the present invention may refer to model hyperparameters that influence the model selection task of the at least one ML algorithm, such as the topology, the size of the neural network, the number of input features, or the manner in which training labels are processed during model training.

[0035] By "predetermined" it is meant that at least one hyperparameter must be fixed before carrying out the method of the present disclosure. The selection of hyperparameters constitutes one embodiment of the present invention. The determining step b) simply selects the particular value for the subject from a given set of hyperparameters.

[0036] Several hyperparameters can be chosen, but the number of input features v must be included among them. If only one hyperparameter is fixed, it is the number of input features v.

[0037] In some non-limiting embodiments, the set of predetermined hyperparameters includes an optimistic model constant λ (called the cure rate), where a particular value λ1 is determined for the subject, and wherein the at least one ML algorithm is trained using the subject's residual indicator labels generated from the training indicator labels obtained for the subject by the optimistic model configured by λ1.

[0038] Preferably, there are two pre-defined hyperparameters of the at least one ML algorithm: v, the number of input features, and λ.

[0039] Training indicator labels

[0040] Training index labels refer to assessments of an index of the subject's mental state at different times during the monitoring period, which are used to train the at least one ML algorithm.

[0041] As described above, in some non-limiting embodiments, the set of predetermined hyperparameters includes a cure rate of an optimistic model λ, in which case a specific value λ1 is determined for the subject, and the at least one ML algorithm is trained using the subject's remaining index labels generated from the training index labels by the optimistic model configured by λ1.

[0042] By "residual index" is meant the difference between the training index and the results of the "optimistic model", which predicts the evolution of the index based only on λ1 and previous evaluations of the index.

[0043] The optimistic model can follow the formula:

number

[0044] ML algorithms

[0045] In some non-limiting embodiments, the at least one ML algorithm is configured using another particular value for the subject of the number of input features v1 and / or at least one other predetermined hyperparameter.

[0046] By "configured" is meant arranged in its functional units, in its main characteristics, in a manner to process input data.

[0047] In particular, in some non-limiting embodiments, the at least one ML algorithm is a multilayer perceptron (MLP), and the depth and / or number of neuron units per layer of the multilayer perceptron (MLP) is a function of a particular value v1 for the subject and / or a function of the particular value for the subject of the at least one other predetermined hyperparameter.

[0048] Determining hyperparameter values

[0049] For at least one predetermined hyperparameter v, determining a particular value v1 of v for the subject may include the use of statistical methods and / or machine learning techniques.

[0050] In some non-limiting embodiments, determining the specific value for the subject of the at least one predetermined hyperparameter v, preferably determining the specific values ​​v1 and λ1, comprises at least one optimization process, possibly involving at least one static variable, based on time-dependent physiological characteristics collected from a group of other living subjects or a group of other living subjects and the subject. In non-limiting embodiments where there are multiple predetermined hyperparameters, determining the specific value for the subject may comprise a separate optimization process for each predetermined hyperparameter.

[0051] In some non-limiting embodiments, determining the specific values ​​for the subject of the predetermined one or more hyperparameters is performed by an optimization process of maximizing a metric measuring the accuracy of the prediction and / or minimizing a metric measuring the error of the prediction, wherein the prediction is generated by the at least one ML algorithm configured with the one or more specific values ​​to determine and trained and evaluated using data collected from a group of other living subjects other than the subject.

[0052] Preferably, during the optimization process, the ML algorithm is trained and evaluated using a leave-one-patient-out approach.

[0053] A grid search can be used during the optimization process to select potential hyperparameter values.

[0054] training method

[0055] Another aspect of the invention is a method of training at least one ML algorithm of the prediction method described above to generate a prediction of at least one indicator of a mental state of a living subject p who may suffer from a mental disorder, the training method comprising: generating a training data set using data obtained from the subject p and a group of other living subjects P, said generating including, for each subject: acquiring, during a monitoring period, time-dependent physiological characteristics from at least one acquisition device sensing the subject, optionally together with at least one static variable of the subject; generating training index labels, the labels representing the mental states of the subject at different times during the monitoring period; said generating comprising: For the subject p, for at least one predetermined hyperparameter v corresponding to the number of input features for the at least one ML algorithm, determining a particular value v1 of v based on data collected from a group of other living subjects P or from a group of other living subjects P and the subject p; Using only the training data for subject p, selecting v input features from the acquired physiological features; training the at least one ML algorithm using the training index labels and input features; The present invention relates to the above method, which comprises:

[0056] system

[0057] A further aspect of the present invention relates to a computer-based system for performing the above-described prediction method, comprising a processing unit for determining, for at least one predetermined hyperparameter v corresponding to the number of input features for at least one ML, a specific value v1 of v for the subject, training at least one machine learning algorithm (ML) according to the above-described training method, and generating a prediction of the at least one indicator of the mental state of the subject through the at least one ML algorithm.

[0058] Computer-Based Systems

[0059] By "computer-based" it is meant that the above-described systems and methods according to the present invention can be implemented using any computer system.

[0060] The computer system may comprise any automated system in any hardware and / or software form, such as a laptop, a personal computer, a smartphone, a workstation, a computer terminal, one or more network computers, or any other data processing system or user device.

[0061] Preferably, the computer system comprises at least one processing unit that executes computer program instructions to control the overall operations associated with the methods of the present invention, which may comprise a general-purpose or special-purpose microprocessor, a microcontroller, or both, or other types of CPUs, GPUs, FPGAs, quantum processors, and any combination thereof.

[0062] The computer system may include one or more storage media on which computer program instructions and / or data are stored, which may include magnetic disks, magneto-optical disks, optical disks, flash memory devices, solid state drives (SSDs), cloud storage, or any other type of removable storage medium, or a combination thereof.

[0063] The computer system preferably comprises at least one memory unit that can be used to load instructions and / or data when they are executed by the processing unit.

[0064] The memory unit may include, for example, random-access memory (RAM) and / or read-only memory (ROM).

[0065] The computer system may further include one or more user interfaces, such as any type of display, keyboard, pointing device (e.g., a mouse or trackpad), audio input device (e.g., a speaker or microphone), or any other type of device that allows user interaction with the computer system.

[0066] The computer system may include one or more network interfaces for communicating with other devices.

[0067] User Interface

[0068] In certain non-limiting embodiments, the systems of the present invention may further comprise a user interface for informing the user via images and / or text of the generated prediction of at least one indicator of the subject's mental state and / or selected input features forming the subject's biosignature.

[0069] As used herein, the term "user" encompasses, for example, a person or entity that owns a user computer, e.g., a computing device or wireless device; a person or entity that operates or utilizes a user computer; or a person or entity that is associated in any way with a user computer.

[0070] The user may be a different person than the living subject. The user may be a decision maker of various natures. Preferably, the user is a clinician and the living subject is a patient of the clinician.

[0071] Acquisition Device

[0072] The system of the present invention may include an acquisition device that senses the subject to generate a time-dependent signature.

[0073] In some embodiments, such acquisition devices include wearable devices, such as wrist monitors, smart watches, smart rings, smart collars, smart belts, and / or smart patches.

[0074] The acquisition device may also include other types of smart devices, such as a smart bed, a smartphone, or any other connected device suitable for sensing the subject.

[0075] Preferably, the acquisition device comprises at least one of an accelerometer, photoplethysmography (PPG), an optical sensor, a heart rate detector (e.g., by optical or electrical detection), and an electrodermal activity sensor (e.g., by measuring dermal electrical impedance or skin capacitance) or an electromagnetic sensor, e.g., a UWB position sensor.

[0076] For a more complete understanding of the present invention, a description of some examples will now be given with reference to the accompanying drawings, in which: [Brief explanation of the drawings]

[0077] [Figure 1] FIG. 1 shows the different steps of an example of a prediction method according to the invention. [Figure 2] FIG. 2 illustrates an example of the temporal evolution of MADRS scores recorded for a subject with MDD. [Figure 3] FIG. 3 illustrates part of the training process of an ML algorithm for generating predictions of indicators of a subject's mental state according to the present invention. [Figure 4] FIG. 4 illustrates the different steps of an example decision process for selecting specific values ​​for a subject for a set of hyperparameters. [Figure 5] Figure 5 reports in tabular form the specific values ​​of one set of hyperparameters (λ,ν) chosen for 26 different patients, where λ is the cure rate of the optimistic model and ν is the number of input features. [Figure 6] Figure 6 depicts the origin of the input features selected by two different patients to generate predictions. DETAILED DESCRIPTION OF THE INVENTION

[0078] 1 to 4 are used to describe the various steps of an example method for prediction using at least one ML algorithm, and an example method for training an ML algorithm according to the present invention.

[0079] FIG. 1 shows the four steps of an example of a prediction method according to the present invention.

[0080] Step 1 involves obtaining a set of time-dependent physiological characteristics of the subject during a monitoring period and recording the corresponding index labels assessed at different time points during the monitoring period (Figure 2).

[0081] Step 2 involves determining at least one predetermined hyperparameter v corresponding to the number of input features for the at least one ML algorithm and a particular value v1 of v for the subject using machine learning and / or statistical techniques (FIG. 4).

[0082] Step 3 involves selecting v1 input features from the obtained physiological features using statistical methods and / or machine learning techniques (Figure 3).

[0083] The final step involves using at least one ML algorithm to generate a prediction of the at least one indicator of the mental state of the subject based on the selected v1 input features.

[0084] FIG. 2 shows an example of a record of indicator labels for a subject whose psychiatric disorder includes MDD and whose indicator is a MADRS score.

[0085] Point 1 in Figure 2 represents the clinical assessment of the subject's mental state performed by a clinician during monthly visits over a 6-month period. These clinical assessments constitute the index label.

[0086] Line 2 is the real evolution of the subject's MADRS score, unknown except for the date of the clinical visit.

[0087] The clinical assessment performed on day d of the monthly clinical visit can be extended within a confidence interval centered on the visit date, depending on the type of labeled data and the periodicity of the labels. In 3, this extension period is selected to be 5 days before and 5 days after the clinical visit. This extension approach can be justified by considering the test-retest reliability of MADRS scores over several days, as described in the literature. It primarily serves to generate more training index labels for training the ML algorithm.

[0088] Line 4 is the output of the optimistic model, which estimates the MADRS evolution based only on the most recent clinical visit and a constant λ (called the cure rate). On any given day, the model predicts the MADRS value given the rate of improvement based on the clinician's past clinical assessment of the patient. The difference between the actual MADRS value and the MADRS value predicted by this optimistic model is called the residual MADRS and is shown in 5.

[0089] The six months can be divided into a training period 6, in which the obtained index labels are used to train the ML algorithm, and an evaluation period 7, in which the index labels are used to evaluate the ML algorithm.

[0090] FIG. 3 illustrates the third and final part of the training process of 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 present invention, where the predetermined hyperparameter is a pair (λ,ν), where λ is the cure rate of an optimistic model and ν is the number of input features for the at least one ML algorithm.

[0091] During the first part of the training process shown in Figure 2, the following dataset was collected over a 6-month period for subject p: One set of physiological characteristics was acquired each day during the period. The index label was assessed at the first visit on day d0 and at monthly visits d1, d2, d3, d4, d5, and d6 for 6 months, with assessments extending up to 5 days before and after each visit.

[0092] During the second part of the training process (described below in Figure 4), the specific values ​​(λ1,ν1) for subject p of the (λ,ν) pair were determined.

[0093] At the start of the third and final step, the dataset acquired for subject p is preprocessed: data from the first three months are used as the training / validation dataset, and the remaining data are used as the test set.

number

[0094] In the selection process S, the feature set X Train / val Among them, label Y using the Hilbert-Schmidt independence criterion (HSIC) Train / val The ν1 features most correlated with are selected and the new input matrix

number

[0095] Metric Label Y Train is input to the optimistic model O constructed by λ1 in the following equation:

number

[0096] Selected features below

number

[0097] The parameters chosen for MLP training were: a batch size of 16, 500 training epochs, and a 5-epoch early stopping callback to monitor improvement in validation loss. The model was trained on the train dataset, as defined above, and validated on the validation dataset. This process was repeated 11 times to smooth random fluctuations due to kernel initialization and to avoid inaccurate predictions due to potential local minima in the model's parameter space, and the final output prediction was set to the median of the predictions.

[0098] The output of the MLP is summed with the base index label B to give the full MADRS score prediction I (FIG. 3).

[0099] The optimistic model constructed by λ1, the one feature selection process, and the MLP constructed by v1 form the following specific model called Signature Based Model of Depression (SiBaMoD) for patient p:

number

[0100] FIG. 4 illustrates at 10 the step of determining specific values ​​(λ1,ν1) of given hyperparameters (λ,ν) for subject p using an optimization process in an exemplary embodiment of the present invention.

number

[0101] As shown in Figure 2, the following data sets were collected over a 6-month period for a group of other living subjects P than subject p. For each subject: One set of physiological characteristics was acquired each day during the period. The index label was assessed at the first visit on day d0 and at monthly visits d1, d2, d3, d4, d5, and d6 for 6 months, with assessments extending up to 5 days before and after each visit.

[0102]

number

number

number

[0103] 5 is a table reporting an example of a set of specific values ​​selected for 26 subjects in one exemplary embodiment of the present invention, where SiBaMoD(λ,ν) is the psychiatric disorder MDD and the indicator is the MADRS score. Most patients selected the same set of values, but four patients selected different sets.

[0104] Figure 6 illustrates the origin of features selected by two different patients to form their signatures. The circular histograms represent the importance of each feature group in the patient's signature, weighted by the number of features in that category. In Patient 1, features describing disturbances in sleep are most prevalent, while in Patient 5, features describing respiratory rate are most prevalent. As can be seen from these examples, biosignatures capture the heterogeneity of depression expression at the individual level and therefore have the potential to direct treatment in a more personalized manner.

[0105] The invention is not limited to the examples described above.

[0106] Other types and numbers of predetermined hyperparameters may be selected. Determining specific values ​​for a subject may rely on statistical methods instead of using the ML algorithm in the prediction process. Static variables may be collected from the subject along with physiological characteristics and used during the determination process.

Claims

1. 1. A method of generating a prediction of at least one indicator of a mental state of a living subject at risk of suffering from a mental disorder using at least one machine learning algorithm (ML), comprising: a. acquiring at least one set of time-dependent physiological characteristics of said subject, optionally together with at least one static variable of said subject; b. For at least one predetermined hyperparameter v corresponding to the number of input features for the at least one machine learning (ML) algorithm, a particular value v of v for the subject 1 determining the c. v of the acquired time-dependent physiological features for the at least one ML algorithm 1 selecting input features, 1 the input features forming a biosignature of the subject; d. using the at least one machine learning (ML) algorithm to generate a prediction of the at least one indicator of the mental state of the subject. The method comprising:

2. The selected v 1 The features are identified as v that are most correlated with the training index labels obtained for the subject and used to train the at least one ML algorithm, preferably using the Hilbert-Schmidt independence criterion (HSIC). 1 The method of claim 1 , wherein the individual features are:

3. The predetermined hyperparameters include a constant λ (called the cure rate) of the optimistic model, where a particular value λ 1 is determined for the subject, wherein the at least one ML algorithm determines λ 1 3. The method of claim 1, wherein the optimistic model is trained using remaining index labels of the subject generated from the training index labels obtained for the subject by the optimistic model constructed by:

4. The method of claim 3 , wherein the predetermined hyperparameters are λ and ν.

5. The optimistic model is of the form [Equation 1] 【number】 The method according to claim 3 or 4.

6. The at least one ML algorithm may be implemented using a number of input features, v 1 and / or using another specific value for the subject of at least one other predetermined hyperparameter.

7. The at least one ML algorithm is a multilayer perceptron (MLP), and the depth and / or number of neuron units per layer of the MLP is set to a specific value v 1 and / or is a function of the particular value for the subject of the at least one other predetermined hyperparameter.

8. 8. The method of claim 1, wherein determining the specific values ​​for the subject of the predetermined hyperparameter(s) is performed by an optimization process of maximizing a metric measuring the accuracy of the prediction and / or minimizing a metric measuring the error of the prediction, wherein the prediction is generated by the at least one ML algorithm configured to determine the specific values ​​and that has been trained and evaluated using data collected from a group of other living subjects other than the subject.

9. 9. The method of claim 8, wherein the at least one ML algorithm is trained and evaluated using data collected from a group of other living subjects using a leave-one-patient-out approach.

10. 10. The method of claim 8 or 9, wherein a grid search is used during the optimization process to select potential hyperparameter values.

11. 11. The method of any one of claims 1 to 10, wherein the psychiatric disorder is Major Depressive Disorder (MDD) and the at least one indicator is the Montgomery-Asberg Depression Rating Scale (MADRS).

12. The method according to any one of claims 1 to 11, wherein the physiological features comprise features representative of physical activity and / or heart rate and / or heart rate variability and / or respiratory rate and / or sleep, preferably all of them.

13. 13. A method of training the at least one ML algorithm of the method of any one of claims 1 to 12 to generate a prediction of at least one indicator of a mental state of a living subject p who may suffer from a mental disorder, said training method comprising: generating a training data set using data obtained from the subject p and a group of other living subjects P, wherein said generating comprises, for each subject: acquiring, during a monitoring period, time-dependent physiological characteristics from at least one acquisition device sensing the subject, optionally together with at least one static variable of the subject; generating training index labels, the labels representing the mental states of the subject at different times during the monitoring period; said generating comprising: For the subject p, for at least one predetermined hyperparameter v corresponding to the number of input features for the ML algorithm, a particular value v of v is determined based on data collected from a group of other living subjects P or from a group of other living subjects P and the subject p. 1 determining the using only the training data for subject p, Among the physiological characteristics obtained, v 1 selecting input features; training the at least one ML algorithm using the training indicator labels and input features; The method comprising:

14. A computer-based system for carrying out the method of any one of claims 1 to 12, comprising: For at least one predetermined hyperparameter v corresponding to the number of input features for at least one ML, a particular value v of v for the subject 1 determining the Training at least one ML algorithm according to 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. The system further comprises a processing unit that executes a program for:

15. 15. The system of claim 14, wherein the system comprises a user interface for informing a user, preferably a clinician, via images and / or text, of the generated prediction and / or the input features selected to generate the prediction.

16. 16. The system of claim 14 or 15, wherein the system comprises an acquisition device that senses the subject during a monitoring period to acquire a set of time-dependent physiological characteristics.

17. 17. The system of claim 16, wherein the acquisition device includes at least one of a photoplethysmograph (PPG), an accelerometer, and an electrodermal activity sensor.

Images