Diabetes monitoring via foot pressure analysis

Abstract

A method performed by a computer system for monitoring diabetes and comprising: a model explainability phase, comprising steps of: - collecting specific foot data from at least one sensor disposed at least within a shoe insole of a user, - collecting a plurality of trained machine learning models configured for decision-making in diabetes diagnosis based on a plurality of collected data, - based on the plurality of trained machine learning models, generating a plurality of specific machine learning models, said specific machine learning models being specified for decision-making in diabetes diagnosis based on the specific foot data, and a model reliability assessment phase, comprising steps of: - evaluating a reliability of each of the specific machine learning model based on neutrosophic sets and intuitionistic fuzzy sets, - based on the result of the evaluation step, selecting the optimized machine learning model for monitoring diabetes based on the specific foot data.

Description

Description Title: Diabetes monitoring via foot pressure analysis Technical Field

[0001] This disclosure pertains to the field of diabetes monitoring based on sensor data and moreparticularly to diabetes diagnosis using artificial intelligence.Background Art

[0002] The diagnosis and continuous monitoring of diabetes represent a complex medical challenge,notably due to multi-contributing factors, plurality of diabetes classifications and symptoms and multi-level complications risks.

[0003] Thus, diabetes monitoring usually requires a frequent professional supervision to provideensure a careful analysis of the evolving symptoms and medical record data of the patients. A heavymonitoring routine for both the patient and the healthcare practitioners is required.

[0004] Technologies related to diabetes diagnosis involving sensor technology and artificialintelligence exist. Yet, most technologies still raise issues in diagnosis accuracy, computationheaviness and / or invasiveness and burden for the patients for example. Indeed, due to the mutli-level factors, symptoms and risks of diabetic conditions, monitoring diabetes usually requires a largeand diversified amount of data, gathered via different monitoring wearables (e.g., glucose monitors,diabetic foot monitor…). Such data gathering often is invasive and burdening for the patients.Moreover, the processing and analysis of such gathered data often require data structuring, which isresource-intensive during computation. There is also a complexity in obtaining accurate and reliablediagnosis results, due to the technical limitations of the sensors, data and algorithmic models used.

[0005] Such issues overall lead to a difficulty in early detecting and monitoring diabetes in anautonomous, proactive and practical way for the individuals. The need for a healthcare practitionerto systematically interpret and assess information for a reliable diabetes diagnosis remains present.

[0006] As a consequence, there is a need to lighten the healthcare burden for both individuals andcaregivers in monitoring diabetes, while improving both the accuracy and reliability of diabetesdiagnoses. Summary

[0007] This disclosure improves the situation.

[0008] It is proposed a method performed by a computer system for monitoring diabetes, the methodcomprising: a model explainability phase, comprising steps of: -collecting specific foot data from at least one sensor disposed at least within a shoe insoleof a user,- collecting a plurality of trained machine learning models configured for decision-making indiabetes diagnosis based on a plurality of collected data, -based on at least the plurality of trained machine learning models, generating a plurality ofspecific machine learning models, said specific machine learning models being specified for decision-making in diabetes diagnosis based on the specific foot data, and a model reliability assessment phase, comprising steps of: -evaluating a reliability of each of the specific machine learning model based on neutrosophicsets and intuitionistic fuzzy sets, -based on the result of the evaluation step, selecting the optimized machine learning modelfor monitoring diabetes based on the specific foot data.

[0009] Advantageously, the proposed method enables to provide a diabetes monitoring anddiagnostic by using a simple and affordable sensor system, simply disposed within the shoe insoleof the system user. Thus, the proposed disclosure reduces the financial burden and the technicalcomplexity on patients, while enabling to provide an immediate, real-time or almost real-timefeedback, which is time-savvy compared to traditional methods for diabetes diagnosis, which require medical appointments and waiting periods for diabetes results. Based on the proposed method, patients can monitor their diabetes condition without reliance on caregivers or medical professionals. Plus, the proposed method relies on at least one sensor which can be integrated into any footwear,making such proposed method widely accessible. Such accessibility of use advantageously enablescontinuous or frequent monitoring, which enables early detection of diabetic evolution, potentially improving health outcomes of the user. The present disclosure enhances patient independence and proactive health management for diabetes monitoring.

[0010] To that end, it is proposed a multiple stage learning and optimization framework fordetermining an optimized machine learning model for diabetes diagnosis and monitoring.Advantageously, such framework includes an unsupervised two-stage framework, via both the modelexplainability phase and the model reliability assessment phase.

[0011] Advantageously, the model explainability phase as proposed synergistically combines aplurality of trained machine learning models to enhance the predictive performance of the optimizedmachine learning model. Moreover, the proposed method relies on data input corresponding to collected specific foot data, which may be unstructured. Thus, the proposed framework advantageously enables to extract patterns and knowledge from unstructured data collected fromsensor data streams, without the necessity of manual labeling. Indeed, the proposed methodadvantageously relies on unsupervised learning methodologies, which enable to identify correlationswithin potentially multi-dimensional data spaces involved in the processed plurality of trainedmachine learning models.

[0012] Advantageously, the model reliability assessment phase as proposed employs a hybridframework combining both intuitionistic fuzzy logic and neutrosophic sets to delineate the outputs of the unsupervised learning framework of the model explainability phase, into calibrated and controlledreliability assessment. Such phase advantageously relies on fuzzy set theory to interpretunsupervised outputs as probabilistic diagnoses rather than deterministic ones. The proposedmethod relies on both fuzzy inference system with medical heuristics and assessment to evaluate the degree of membership of each model output.

[0013] The proposed thus enables to provide an optimized machine learning model, adapted andbuilt to predict based on specific foot data collected from the user as data input.

[0014] By a computer system, it may be understood a diabetes monitoring system or part of suchdiabetes monitoring system. The computer system includes at least one sensor disposed by the user and a processing unit or module configured to perform the proposed method.

[0015] By specific foot data, it may be understood one specific type of data collected via at least onesensor disposed near the foot area of the user. In particular, such specific foot data corresponds toonly one type of data, for example pressure data. Such specific foot data may be collected as manytimes as the user wears the sensor and uses the computer system for monitoring diabetes. Suchspecific foot data may thus be iteratively, periodically, punctually or continuously collected, thusrunning iterations of the proposed method. In particular, the specific foot data may correspond tounlabeled and unstructured data.

[0016] By a plurality of trained machine learning models, it may be understood machine learningmodels resulting from a supervised learning stage for decision-making in diabetes diagnosis. Inparticular, such supervised learning stage includes a training step and may include a validation step and a test step resulting in the trained machine learning models. Each of such trained machinelearning models notably relies on the plurality of collected data as data inputs. By a plurality ofcollected data, it may be understood different and various types of data of different natures. Forexample, such plurality of collected data may include temperature data, pressure data, humidity data,medical information data. Such collected data serves as datasets for the learning process of thetrained machine learning models. In particular, such collected data are labeled and structured toserve as data input for the initial machine learning models to be trained. The trained machine learningmodels may have different parameters, hyperparameters, specifications, rules etc.

[0017] By a plurality of specific machine learning models, it may be understood machine learningmodels resulting from an unsupervised learning stage during the model explainability phase. Inparticular, such unsupervised learning stage does not require training on the collected specific footdata. The specific machine learning models are generated from the trained machine learning modelsSuch generating may use a combination of several machine learning algorithms to combine andenhance the respective performances of the plurality of trained machine learning models. Inparticular, the specific machine learning models are generated so as to take only the collected specific foot data as data input. In particular, the data input of the specific machine learning modelsdiffers from the data inputs of the trained machine learning models.

[0018] By a model explainability phase, it may be understood generating data related to a pluralityof specific machine learning models based on an unsupervised learning framework and solely relyingon the specific foot data. In particular, such model explainability phase includes providing specificmachine learning models with explainable aspects and enhanced performance for diabetes diagnosis with respect to the trained machine learning models.

[0019] By evaluating a reliability of a specific machine learning model, it may be understooddetermining the trustworthiness or untrustworthiness of the specific machine learning model in thedecision-making for diabetes diagnosis, based on the collected specific foot data of the user. Suchevaluation notably takes into account the uncertainty and indeterminacy lying in the parameters, hyperparameters, model structures and data input of the specific machine learning models, as well as additional medical knowledge in case of insufficient means to assess the model reliability, usingfuzzy set logic processes.

[0020] The following features of the method may optionally be performed, either separately or incombination with one another:

[0021] In an embodiment of the present disclosure, the specific foot data corresponds to footpressure data collected from at least one pressure sensor disposed within the shoe insole of the user.

[0022] In an embodiment of the present disclosure, the method further comprises:a modeling phase, comprising steps of: -collecting data including sensor-based data collected from a plurality of sensors disposed atleast within a shoe insole of a plurality of individuals, and medical information related to diabetes diagnosis, -based on the collected data, determining the plurality of trained machine learning models.

[0023] As a consequence, the proposed method advantageously enables to determine the pluralityof trained machine learning models, which further serve as base models to generate the specificmachine learning models and select the optimized machine learning model. In particular, suchmodeling phase may advantageously be performed beforehand with respect to the modelexplainability phase and the model reliability assessment phase, and by a distinct processing device. The computational complexity of determining such plurality trained machine learning models does not impact the proposed method for diabetes diagnosis, as the computer system for monitoringdiabetes in under use by the user.

[0024] Advantageously, the proposed method only requires performing such modeling phase once,in order to obtain the plurality of trained machine learning models. Only the unsupervised learningstage as proposed can then be iterated as many times as specific foot data is collected by the user.

[0025] The modeling phase may advantageously serve as a proof of concept of the diabetesmonitoring method that the trained plurality of machine learning models further used in theunsupervised learning stage have the prediction capabilities to address the diabetes diagnosis ormonitoring problem.

[0026] In an embodiment of the present disclosure, determining the plurality of trained machinelearning models further comprises a step of:preprocessing the collected data, said preprocessing including steps of cleaning,normalizing, feature extracting and / or feature selecting.

[0027] Advantageously, such pre-processing notably provides the machine learning models to betrained with structured and labeled data for the model training. Notably, the collected medicalinformation may enable to label the collected data.

[0028] In an embodiment of the present disclosure, determining the plurality of trained machinelearning models further comprises steps of: selecting a plurality of initial machine learning models, and for each selected initial machine learning model: splitting the collected data into at least a training dataset and a test dataset, training the selected initial machine learning model using the training dataset, evaluating a performance of the selected initial machine learning model after the training step, using the test dataset, depending on the performance evaluation step, adjusting parameters and / or hyperparameters of the selected predefined machine learning model, the trained machine learning model resulting from said adjustment step.

[0029] In an embodiment of the present disclosure, the modeling phase further comprises:collecting updated data including updated sensor-based data and / or updated medical information, performing a continuous learning process based on said collected updated data.

[0030] Advantageously, the plurality of trained machine learning models may be improved asupdated collected data is provided to the trained machine learning models for continuous learning. Such continuous learning then enables to further enhance the performance of the optimized machine learning model which is generated and selected based on the plurality of trained machine learning models.

[0031] In an embodiment of the present disclosure, the modeling phase is based on supervisedlearning processes using at least the collected data.

[0032] As a consequence, the modeling phase may be functionally and temporally distinguishedfrom the unsupervised learning processes of the model explainability phase and the model reliability phase. The proposed method thus relies on both the supervised learning stage and the unsupervisedlearning stage, which enables to make use of the prediction performances of the trained machinelearning models to interpret unlabeled, unstructured, and real-time data input formed by the specificfoot data.

[0033] In an embodiment of the present disclosure, the collected data includes sensor-based data,said sensor-based data including at least one element among:- foot humidity data collected on at least one humidity sensor disposed within the shoe insoleof a plurality of individuals, -foot temperature data collected on at least one temperature sensor disposed within the shoeinsole of the plurality of individuals, and -foot pressure data collected on at least one pressure sensor disposed within the shoe insoleof the plurality of individuals.

[0034] In an embodiment of the present disclosure, the foot humidity data includes at least oneelement among: -ambient humidity data, related to a moisture level of the air surrounding the humidity sensor,- insole humidity data, related to a moisture level within the shoe insole of each of the pluralityof individuals, -humidity variation data, related to variations of a humidity level over a predefined period, and- humidity fluctuation data, related to paces of changes of the humidity level.

[0035] In an embodiment of the present disclosure, the foot temperature data includes at least oneelement among: -ambient temperature data, related to a temperature of the air surrounding the temperaturesensor, -foot surface temperature data, related to a temperature at a surface of a foot of each of theplurality of individuals, -temperature variation data, related to variations of a temperature over a predefined period,and -temperature fluctuation data, related to paces of changes of temperature.

[0036] In an embodiment of the present disclosure, the foot pressure data includes at least oneelement among: -plantar pressure data, related to a pressure level exerted on different plantar regions of afoot of one among the user and the plurality of individuals, -pressure distribution data, related to a mapping of pressure points on the foot of one amongthe user and the plurality of individuals, -pressure variation data, related to variations of the pressure level over a predefined period,- pressure variation pattern data, related to patterns of pressure changes over predefinedtimes, and -peak pressure data, related to extremal pressure values exerted in the different plantarregions of the foot of one among the user and the plurality of individuals.

[0037] In an embodiment of the present disclosure, the medical information related to diabetesdiagnosis includes at least one element among: -predefined glycemic level thresholds,- predefined diabetes diagnosis criteria,- predefined metrics for assessing diabetes complication risks,- predefined diabetes treatment protocols for different diabetes stages,- predefined lists of symptoms associated to different diabetes stages,- lists of diabetes classification types,- values of blood glucose test levels,- glycated hemoglobin – HbA1c – average values,- insulin efficiency values,- risk level values for diabetes complication risks,- effectiveness values of treatment protocols.

[0038] In an embodiment of the present disclosure, the model explainability phase is based onunsupervised learning processes.

[0039] In an embodiment of the present disclosure, the model explainability phase is performed usingat least one element among Generative Adversarial Networks – GAN -, random forest algorithms,pattern recognition algorithms.

[0040] In an embodiment of the present disclosure, the model explainability phase further includesa step of: -performing a feature importance analysis on the plurality of trained machine learning modelsbased on the specific foot data.

[0041] In an embodiment of the present disclosure, the model explainability phase further includesa step of: -performing a transfer learning on the plurality of trained machine learning models, in orderto adapt the plurality of trained machine learning models for decision-making in diabetes diagnosis based on the specific foot data as an only data input.

[0042] In an embodiment of the present disclosure, the model explainability phase further includesa step of: -performing a hyperparameter tuning of each of the plurality of trained machine learningmodels, in order to adapt a learning process of each of the plurality of trained machine learning model to learn with the specific foot data as an only data input.

[0043] In an embodiment of the present disclosure, the model explainability phase further includesa step of: -performing a cross-validation to evaluate a model performance of each of the plurality oftrained machine learning models, using the specific foot data as an only data input.

[0044] In an embodiment of the present disclosure, the model explainability phase further includesa step of: -determining explainability parameters related to an explainability of each of the plurality oftrained machine learning models.

[0045] In such embodiment, the explainability parameters may include at least one element among:a Root Mean Square Error – RMSE- value, a Mean Absolute Error – MAE- value, an accuracy value,an F1-score value.

[0046] Advantageously, the proposed method operates within the paradigm of explainable artificialintelligence (or xAI), which offers transparency into the decision-making process, which is a crucial requirement for healthcare applications.

[0047] In an embodiment of the present disclosure, the model explainability phase further includesa step of: -determining a meta-model of the plurality of specific machine learning models based on acomparison of a performance of each of the plurality of specific machine learning models with a predefined baseline model.

[0048] Advantageously, the method proposes to determine an optimized prediction tool whichcombines the best features of the plurality of specific ML models, in order to obtain an enhanced predictive performance for diabetes diagnosis. In particular, the generated meta-model provides a better predictive performance with respect to each of the specific machine learning models.

[0049] In an embodiment of the present disclosure, the model explainability phase further includesa step of: -generating model report data related to the plurality of specific machine learning modelsand / or to a meta-model generated based on such plurality of specific machine learning models.

[0050] As a consequence, the model report data advantageously provides all input parameters,variables, hyperparameters and the used algorithms enabling to define and shape a meta-model for decision-making in diabetes diagnosis. Data contained in such model report data may then advantageously be assessed or evaluated in order to determine a reliability level of the resulting ML model to be used for diabetes diagnosis.

[0051] In an embodiment of the present disclosure, the model explainability phase further includessteps of:- collecting updated specific foot data,- performing a continuous learning based on said collected updated specific foot data.

[0052] Advantageously, the proposed method dynamically refines and adapts its predictive modelsthrough continuous learning, thus ensuring that the generated predictions are updated and relevant. As a consequence, the unsupervised learning stage may advantageously improve itself via continuous learning, notably without requiring to reiterate the modeling phase.

[0053] In an embodiment of the present disclosure, the model reliability assessment phase furthercomprises steps of: -integrating neutrosophic sets to the specific machine learning models,- integrating intuitionistic fuzzy sets to the specific machine learning models,- based on the integrated neutrosophic sets and intuitionistic fuzzy sets, determiningmembership values associated to outputs of each of the specific machine learning models, -establishing assessment thresholds with respect to the membership values, saidassessment thresholds including at least a high assessment threshold and a low assessment threshold.

[0054] As a consequence, the method proposes to introduce uncertainty and indeterminacy aspectsin the model outputs, based on fuzzy set logic. The proposed method thus advantageously enablesto assess the reliability of the generated specific machine learning models (or equivalently, of the generated meta-model resulting from such specific ML models) based on controlled criteria,thresholds and takes into account the stochastic and complexity constraints of machine learningmodels for decision-making in diabetes diagnosis.

[0055] In particular, by defining assessment thresholds at the scale of outputs of each of the specificmachine learning models, the proposed method enables to take into account the diversity and differences between the plurality of machine learning models processed throughout the proposedmethod. The reliability assessment of each specific machine learning model thus takes into accountthe correlations and inherent structures forming such specific machine learning model, thus enhancing the relevance and the trustworthiness of such assessment.

[0056] In an embodiment of the present disclosure, the model reliability assessment phase furthercomprises steps of: for each specific machine learning model, -evaluating the specific machine learning model based on the assessment thresholds,- if a membership value of an output of the specific machine learning model exceeds the highassessment threshold, evaluating the specific machine learning model as having a highreliability level, -if the membership value of the output of the specific machine learning model is below thelow assessment threshold, evaluating the specific machine learning model as having a lowreliability level,- if the membership value of the output of the specific machine learning model is between thehigh assessment threshold and the low assessment threshold, evaluating the specificmachine learning model as having a medium reliability level.

[0057] Advantageously, the proposed method enables to associate each specific machine learningmodel with a reliability level, such reliability level being directly linked to a membership functionleveraging and to model output interpretation.

[0058] By comparing the membership value of an output of a specific machine learning model withan assessment threshold, it is understood comparing the one or several membership values associated to the output with the assessment threshold. By an output exceeding or being below an assessment threshold, it may be understood that at least one, all or a predefined range of number of membership values exceed the assessment threshold.

[0059] In an embodiment of the present disclosure, the model reliability assessment phase furthercomprises steps of: -transmitting the specific machine learning models having a medium reliability level to acomplementary model assessment unit, the complementary model assessment unit being configured to evaluate the transmitted specific machine learning models based on a medical assessment threshold, -if the output of the transmitted specific machine learning model is below the medicalassessment threshold, transferring said transmitted specific machine learning model to a model storage storing the plurality of trained machine learning models,- if the output of the transmitted specific machine learning model is above the medicalassessment threshold, evaluating the specific machine learning model as having a high reliability level.

[0060] Advantageously, the proposed method enables to sort trustworthy results from untrustworthyresults. As trustworthy results may be exploited, for example by triggering patient notification, untrustworthy results invoke a robotic process automation protocol to solicit medical expert intervention for validation or rejection. The proposed method thus provides a multi-level reliabilityassessment of the diagnosis model, thus enhancing its reliability and trustworthiness. The proposedapproach leverages fuzzy decision-making, thus enhancing interpretability and controlled flexibility in the diabetes diagnostic process. The proposed method also incorporates an element of automated decision support, via the complementary model assessment unit, in order to manage the variability inherent in diabetes prognostication.

[0061] Advantageously, the method proposes to delegate the assessment task to an additionalassessment unit for further and supplementary assessment. Such additional assessment is advantageously performed only when necessary, that is when the fuzzy set logic processes do nothave sufficient data knowledge to assess the reliability of a specific machine learning model. Theproposed method thus adds flexibility and complementary analysis when required.

[0062] In an embodiment, the model reliability assessment phase is performed on a meta-modelgenerated based on the plurality of specific ML models. In such embodiment, the model reliabilityassessment phase includes steps of evaluating a reliability of the meta-model based on neutrosophic sets and intuitionistic fuzzy sets, and based on the result of the evaluation step, determining the optimized machine learning model for monitoring diabetes based on the specific foot data.

[0063] In an embodiment of the present disclosure, the optimized machine learning model is selectedamong the specific machine learning models having a high reliability level.

[0064] As a consequence, the proposed method ensures an enhanced and optimized performanceand accuracy of the provided machine learning model, by selecting the optimized machine learningmodel based on satisfactory and reliable metrics.

[0065] In an embodiment of the present disclosure, the method further comprises a diabetesmonitoring phase including a step of: -transmitting a diabetes monitoring result to a final user interface, said diabetes monitoringresult being determined based on an output of the optimized machine learning model based on the specific foot data.

[0066] As a consequence, the proposed method enables to provide a user with a diabetes diagnosisresult (or equivalently, with a diabetes monitoring result), which accuracy and reliability are optimized. The user is thus advantageously provided with a trustworthy and accurate result regarding his or herdiabetes diagnosis. Moreover, the proposed method enables to obtain such diabetes result in asimple way, by simply using the sensor to collect the specific foot data. The diagnosis result may beobtained in a simple, practical, nearly instantaneous way, since neither data labeling andstructuration phase, nor training phase is required.

[0067] In an embodiment of the present disclosure, the final user interface may be at least oneelement among: -a final user wearable,- a final user mobile device,- a human-machine interface,- a cloud interface.

[0068] Advantageously, the proposed method enables a user to obtain a diabetes monitoring ordiagnosis result by its own means, based on simple, non-intrusive and practical devices, with no orseldom medical intervention. The proposed method thus enables a more frequent, even continuousmonitoring of diabetes, which enables a better healthcare progression monitoring and a decreasedburden on the medical practitioners, while ensuring a medical and technical accuracy and reliability of the provided diabetes result.

[0069] According to another aspect of the disclosure, it is proposed a diabetes monitoring systemcomprising computer means for performing a method as aforementioned. According to anotheraspect of the disclosure, it is proposed a computer software comprising instructions to implement atleast a part of a method as aforementioned when the software is executed by a processor. Accordingto another aspect of the disclosure, it is proposed a computer-readable non-transient recording medium on which a software is registered to implement a method as aforementioned when thesoftware is executed by a processor.Brief Description of Drawings

[0070] Other features, details and advantages will be shown in the following detailed description andon the figures, on which: Fig.1

[0071] [Fig. 1] is a schematic representation of a diabetes monitoring system according to anembodiment. Fig.2

[0072] [Fig. 2] is a flowchart showing steps of a diabetes monitoring method according to anembodiment. Fig.3

[0073] [Fig.3] is a flowchart showing steps of a modeling phase according to an embodiment.Fig.4

[0074] [Fig. 4] is a flowchart showing steps of a model explainability phase according to anembodiment. Fig.5

[0075] [Fig. 5] is a flowchart showing steps of a model reliability assessment phase according to anembodiment. Description of Embodiments

[0076] It is now referred to figure 1. Figure 1 is a schematic representation of a diabetes monitoringsystem according to an embodiment of the disclosure.

[0077] The diabetes monitoring system includes at least a model optimization device 6, a modelstorage 5, at least one user sensor 1b, a user terminal 8 and an expert terminal 7.

[0078] The model optimization device 6 is configured to determine an optimized machine learningmodel (also referred to as an optimized ML model) for decision-making in a diabetes diagnosis. Inparticular, the model optimization device 6 is configured to determine such optimized ML modelbased on data accessed from the model storage 5 and from the at least one user sensor 1b. Themodel optimization device 6 is then configured to process such data in a model explainability phaseand a model reliability assessment phase, which will be further described.

[0079] The model storage 5 is configured to store data accessible to the model optimization device6. In particular, the model storage 5 is configured to store a plurality of trained machine learningmodels (also referred to as trained ML models) to be accessed and processed by the modeloptimization device 6.

[0080] The user sensor 1b is configured to measure data related to a user of the diabetes monitoringsystem 1. The user sensor 1b may be connected to the model optimization device 6 via wirelessmeans so as to transmit the measured data to the model optimization device 6. In particular, the usersensor 1b is disposed at least within a shoe insole of the user. The user sensor 1b may correspondto a sensor configured to measure one specific type of data. In particular, the user sensor 1b may bea pressure sensor. In an embodiment, there may be a plurality of user sensors 1b disposed in a footarea of the user. The user sensors 1b is configured to measure specific foot data of the user. Suchmeasurements may be continuous, periodic or punctual. The specific foot data may be measuredwhen the user is walking at various paces, doing exercise or resting for example. In particular, theuser sensor 1b is a non-invasive sensor and may be integrated in the shoe of the user, so that theuser sensor 1b may collect specific foot data in a continuous way when the user wears the shoe.

[0081] The user terminal 8 is configured to communicate with the model optimization device 6 viawireless means. In particular, the user terminal 8 is able to receive signals from the modeloptimization device 6, such as messages and notifications related to a result of a diabetes diagnosisor monitoring for the user of the user terminal 8. The signals from the model optimization device 6may also relate to the optimized ML model to be used for determining the result of the diabetes diagnosis or monitoring. In particular, such user may correspond to the user wearing the user sensor1b. The user terminal 8 may refer to any device, terminal, IoT object, wearable object of the userconnected to the model optimization device 6 via the wireless means. For example, the user terminal8 may be a mobile phone, a tablet, a computer, a smart watch or any other smart device of the user.The user terminal 8 may be configured to receive regular, punctual, periodic signals related to theresult of a diabetes diagnosis or monitoring of the user.

[0082] The expert terminal 7 is configured to communicate with the model optimization device 6 viawireless means. The expert terminal 7 may be accessed by a medical expert, a practitioner, a doctoror any medical staff or caregiver in charge of accessing a medical record of the user whose diabetesresult is provided by the model optimization device 6. In particular, the expert terminal 7 is able toreceive signals from the model optimization device 6, such as messages and notifications related toa result of a diabetes diagnosis or monitoring for a user, or signals to assess or verify the reliabilityof a given ML model. The expert terminal 7 may refer to any device, terminal, IoT object, platformconnected to the model optimization device 6 via the wireless means. For example, the expertterminal 7 may be a mobile phone, a tablet, a computer, a cloud platform, or any other digital device. The expert terminal 7 may be configured to receive regular, punctual, periodic signals related to the result of a diabetes diagnosis or monitoring of the user. The expert terminal 7 may also store oraccess medical information related at least to diabetes diagnosis detained by a medical knowledgeunit 2. The expert terminal 7 may include at least a memory unit 70 and a processing unit 71.

[0083] In an embodiment, the diabetes monitoring system may also include a plurality of individualsensors 1a, at least one medical knowledge unit 2, a modeling device 3 and a model management device 4.

[0084] The modeling device 3 is configured to determine the plurality of trained ML models byperforming a modeling phase 300, which will be further described. In an embodiment, the modelingdevice 3 is a distant device from the model optimization device 6. The modeling device 3 may be configured to train a plurality of initial machine learning models (also referred to as initial ML models)based on collected data. To that end, the modeling device 3 may include a memory unit 30 and aprocessing unit 31, said processing unit including a plurality of processing modules 31a, 31b, 31c,31d to perform various steps of the modeling phase 300. The data collected by the modeling device3 and used to perform the modeling phase 300 may include sensor-based data collected from a plurality of sensors 1a disposed at least within a shoe insole of a plurality of individuals. In particular,such individuals may not include the user. Such individuals may include diabetic and non-diabeticindividuals, diabetes-diagnosed or undiagnosed individuals. The data collected by the modelingdevice 3 and used to perform the modeling phase 300 may also include medical information relatedto diabetes diagnosis retrieved from at least one medical knowledge unit 2.

[0085] The plurality of sensors 1a may be disposed for one or several predefined time intervals withinthe shoe insole of a plurality of pre-selected individuals. Such plurality of sensors 1a are configuredto measure sensor-based data from such plurality of individuals. In an embodiment, such sensor-based data relate to a foot area of the individuals. Such plurality of sensors 1a may include meansfor measuring or collecting such sensor-based data and communication means for transmitting suchsensor-based data to the modeling device 3. The plurality of sensors 1a may for example includehumidity sensors, temperature sensors and / or pressure sensors.

[0086] The at least one medical knowledge unit 2 may be configured to provide medical informationrelated to diabetes diagnosis. Such medical knowledge unit 2 may for example be medical databasesand platforms provided to and / or by medical experts on diabetes. The medical knowledge unit 2 mayalso include digital tools and devices, medical terminals 7 used by medical experts to provide adiabetes diagnosis or to centralize medical information related to diabetic and / or non-diabeticpatients. The medical knowledge unit 2 is configured with wireless means for transmitting medicalinformation to at least the modeling device 3 and / or to any medical terminal 7.

[0087] The model management device 4 is configured to organize a plurality of trained ML modelsbefore storing such trained ML models in the model storage 5. The model management device 4may be configured to organize trained ML models determined by the modeling device 3 and / orreceived from the model optimization device 6. To that end, the model management device 4 mayinclude at least a processing unit 41 and communication means with the modeling device 3 and themodel optimization device 6.

[0088] It is now referred to figure 2. Figure 2 is a flowchart showing steps of a diabetes monitoringmethod 200. Such method may be performed by a diabetes monitoring system 1 as illustrated onfigure 1 or by part of the diabetes monitoring system 1 as represented on figure 1.

[0089] In an embodiment of the present disclosure, the diabetes monitoring method 200 includessteps 300c, 400, 400a, 500 and 500a. In another embodiment, the diabetes monitoring method 200further includes steps 300, 300a and 300b.

[0090] At a step 300, a modeling phase may be performed in order to determine a plurality of trainedML models for decision-making in diabetes diagnosis. Such modeling phase 300 may be performedby a modeling device 3 as represented on figure 1. In particular, the modeling phase 300 is basedon supervised learning processes, so that the collected data during the modeling phase 300 enables to train, validate and adjust a plurality of ML models to predict expected output classes (e.g., theoutput may be “Diabetes” or “Non-Diabetes”). The modeling phase 300 may involve severalthousands or millions of trained machine learning models. The modeling phase 300 will be furtherdetailed in figure 3.

[0091] At an optional step 300a, the plurality of trained ML models may be exploited for diabetesdiagnosis and / or for medical knowledge updating. For example, one or several of the plurality oftrained ML models may be used by the medical knowledge unit 2 and / or by the expert terminal 7 inorder to provide a diabetes diagnosis result to a plurality of patients. Such plurality of trained MLmodels may also be used by medical knowledge units 2 for medical research purposes, for examplein order to update or improve the medical information related to diabetes diagnosis. Such trained MLmodels may be used by inputting data measured on each patient, so as to obtain a binary output ofeach trained ML model, for example for a “Diabetes” or “Non-Diabetes” classification. In particular,the input data has a same format with respect to the collected data used to train and validate the plurality of trained ML models during the modeling phase.

[0092] At a step 300b, the plurality of trained ML models may be organized by the modelmanagement device 4. For example, the plurality of trained ML models may be organized andclassified depending on the algorithm nature of each of the trained ML models. The modelmanagement device 4 may for example gather trained ML models of the plurality of trained MLmodels which are random forest algorithms, convolutional neural networks -CNN-, or support vectormachines -SVM-. Such step 300b enables to organize and classify the plurality of trained ML modelsbased on the model architecture and / or design.

[0093] At a step 300c, the plurality of trained ML models may be stored in the model storage 5. Suchtrained ML models may notably be stored according to the model organization performed by the model management device 4.

[0094] At a step 400, a model explainability phase is performed, in order to generate a plurality ofspecific ML models for diabetes diagnosis. Such model explainability phase 400 may be performedby the model optimization device 6 as represented on figure 1, more particularly by a first processingmodule 611 of the model optimization device 6. Such model explainability phase 400 will be furtherdetailed in figure 4.

[0095] At a step 400a, data related to the plurality of specific ML models as generated at step 400are stored in a report storage 62. In particular, model parameters and data related to the generatedspecific ML models are stored in the report storage 62. For example, such model parameters anddata may be stored in an XML file stored in the report storage 62.

[0096] At a step 500, a model reliability assessment phase is performed, in order to select anoptimized ML model for decision-making in diabetes diagnosis. Such optimized ML model may notably be selected among the plurality of specific ML models as generated during the modelexplainability phase 400. The model reliability assessment phase 500 may be performed at least bythe model optimization device 6, more particularly by a second processing module 611 of the modeloptimization device 6. Such model reliability assessment phase 500 may also involve stepsperformed by the model optimization device 6 and / or an expert terminal 7. The output of the modelreliability assessment phase 500 is an optimized and reliable ML model to be used for diabetesmonitoring or diagnosis of a user. The model reliability assessment phase 500 will be further detailedin figure 5.

[0097] At a step 500a, a diabetes monitoring result may be determined for the user. At step 500a,the optimized ML model as determined at step 500 may be used to determine a diabetes monitoringresult, based on specific foot data collected from the user during step 400. Such diabetes monitoringresult may be determined by the model optimization device 6, by a user terminal 8 or by any other device. In an embodiment, such diabetes monitoring result may be transmitted to the user terminal 8, for example via a human-machine interface of the user terminal 8. In another embodiment, the diabetes monitoring result may also be transmitted to the expert terminal 7 for further diagnosisassessment or recording by a medical expert. The diabetes monitoring result as determined at step500a may output a classification result among “Diabetes” and “Non-Diabetes”. The diabetesmonitoring result as determined at step 500a may also output a classification result related to aadvancement stage of the diabetes (for example, “Diabetic Stage 1” for an early detection of diabetes,“Diabetic Stage 2” for a classic detection of diabetes and “Diabetes Stage 3” for a late detection of diabetes).

[0098] It is now referred to figure 3. Figure 3 illustrates steps of a modeling phase 300 for determininga plurality of trained ML models.

[0099] At the step 301, a plurality of initial ML models is selected, for example by benchmarking.

[0100] Such plurality of initial ML models may include machine learning models with a plurality ofdifferent architectures, designs, structures. For example, the plurality of initial ML models may includeRandom Forests (RF), Support Vector Machines (SVM), Neural Networks (e.g., CNN), Clustering(e.g., K-Means) and / or any other existing machine learning model algorithm type. The plurality ofinitial ML models may be selected based on predefined ML models (such as GPT, BERT, ResNet,Yolo, SSD and Isolation Forest) having satisfying performances (based on a plurality of performancemetrics such as F1 score, MAE, RMSE and AUC-ROC) for healthcare diagnosis and monitoring.

[0101] At a step 302, sensor-based data is collected from a plurality of sensors associated to aplurality of individuals.

[0102] The plurality of individuals may be a sample of individuals including both diabetic individualsand non-diabetic individuals.

[0103] In particular, the plurality of sensors may be disposed within a shoe insole of the plurality ofindividuals. Such plurality of sensors may notably include humidity sensors, temperature sensorsand pressure sensors. In an embodiment, such diversity of sensors are disposed on each individualof the plurality of individuals, so that foot humidity data, foot temperature data and foot pressure datamay be collected for each individual. The plurality of sensors may be configured to collect suchsensor-based data in a continuous way or during predetermined times, when the plurality ofindividuals adopts different postures, have different walk paces and / or perform different activities(e.g., during a physical effort, during a stroll, during sleeping, during an indoor and / or outdoor activityetc.).

[0104] In an embodiment, the foot humidity data includes at least one element among:- ambient humidity data, related to a moisture level of the air surrounding the humidity sensorof each of the plurality of individuals, -insole humidity data, related to a moisture level within the shoe insole of each of the pluralityof individuals, -humidity variation data, related to variations of a humidity level over a predefined period, and- humidity fluctuation data, related to paces of changes of the humidity level.

[0105] In an embodiment, the foot temperature data includes at least one element among:- ambient temperature data, related to a temperature of the air surrounding the temperaturesensor of each of the plurality of individuals,- foot surface temperature data, related to a temperature at a surface of a foot of each of theplurality of individuals, -temperature variation data, related to variations of a temperature level over a predefinedperiod, and -temperature fluctuation data, related to paces of changes of the temperature level.

[0106] In an embodiment, the foot pressure data includes at least one element among:- plantar pressure data, related to a pressure level exerted on different plantar regions of thefoot of each of the plurality of individuals,- pressure distribution data, related to a mapping of pressure points on the foot of each of theplurality of individuals, -pressure variation data, related to variations of the pressure level over a predefined period,- pressure variation pattern data, related to patterns of pressure changes over predefinedtimes, and- peak pressure data, related to extremal pressure values exerted in the different plantarregions of the foot of each of the plurality of individuals.

[0107] The collected sensor-based data may be timestamped and be associated to different labels.In particular, the collected sensor-based data may include a matching between each piece of data and the diabetic or non-diabetic state of the corresponding individual.

[0108] At a step 303, medical information related to diabetes diagnosis is collected.

[0109] Such medical information may notably be provided by one or several medical knowledge units2, as represented on figure 1. Such medical information may for example include domain expertknowledge about diabetes and about diabetes diagnosis. Such medical information may also includediabetes level taxonomy. In an embodiment, the medical information collected at step 303 mayinclude at least one element among:- predefined glycemic level thresholds,- predefined diabetes diagnosis criteria,- predefined metrics for assessing diabetes complication risks,- predefined diabetes treatment protocols for different diabetes stages,- predefined lists of symptoms associated to different diabetes stages,- lists of diabetes classification types,- values of blood glucose test levels,- glycated hemoglobin – HbA1c – average values,- insulin efficiency values,- risk level values for diabetes complication risks,- effectiveness values of treatment protocols.

[0110] The collected data formed by both the sensor-based data collected at step 302 and themedical information collected at step 303 forms the data input of each of the plurality of initial MLmodels.

[0111] At a step 310, a preprocessing is performed on such data input. Such preprocessing may forexample include steps of data cleaning, normalizing, feature extracting and / or feature selecting onthe data collected at steps 302 and 303. In an embodiment, such step 310 may include a datalabelling with labels associated to “Diabetes” class and “Non-Diabetes” class. In particular, thepreprocessing performed at step 310 results in data input corresponding to structured data. Referringto the Examples section of the present disclosure, [Table 1], [Table 2], [Table 3], [Table 4], [Table 5]and [Table 6] illustrate examples of such structured data used as data input for the plurality of initialML models.

[0112] At a step 320, the data input is partitioned into training and test datasets. In anotherembodiment, the data input may be partitioned into training, validation and test datasets. In particular,the data input partitioning may be performed at step 320 for each initial ML model. In other words,the training and test datasets (and notably the data representation of each class and size within eachdataset) may differ depending on the initial ML model.

[0113] At a step 330, a training of each of the initial ML models is performed using the training dataset.

[0114] Optionally, at a step following step 330 (not represented on figure 3), a validation of each ofthe initial ML model after the training may be performed using the validation dataset.

[0115] At a step 340, a test of each of the initial ML models after training is performed, using the testdataset. Such step 340 enables to evaluate a performance of each of the initial ML model after thetraining step 330 (and optionally after the validation step). In particular, at such step 340, each of thetested initial ML model outputs a classification result, for example as either “Diabetes” or “Non-Diabetes”, from the test dataset input.

[0116] At a step 350, a model adjustment is performed on each of the initial ML models after the teststep 340. Such model adjustment 350 may include adjusting parameters and / or hyperparameters ofeach of the initial ML model based on the prediction accuracy performed on the test dataset duringthe test step 340, in order to enhance the performance and accuracy of such tested initial ML model. After step 450, the model adjustment results in a plurality of trained ML models.

[0117] Optionally, at a step 360, a continuous learning may be performed on the plurality of trainedML models. To that end, updated data may be collected, for example as updated foot data and / orupdated medical information, at further iterations of steps 302 and 303. Such updated data may becollected based on another plurality of individuals (that is, a different sample of individuals) or onupdated data retrieved from the same plurality of individuals (for example, in a different measurementcontext or at another time). Based on the updated data, a continuous learning process may beperformed at step 360, so as to update the trained ML models and thus to improve the predictionperformance of the trained ML models based on updated collected data.

[0118] The modeling phase 300 thus enables to output a plurality of trained ML models, each trainedML model being described by a specific set of rules, parameters, hyperparameters, data structureand providing a given prediction accuracy.

[0119] It is now referred to figure 4. Figure 4 illustrates steps for performing a model explainabilityphase 400. During such model explainability phase 400, a specific ML model is generated for eachtrained ML model stored in the model storage 5. Such trained ML model may for example result fromthe modeling phase 300 as previously described. Such generated plurality of specific ML models isnotably configured to make predictions based solely on specific foot data as data input. Moreover,such generated specific ML models have an enhanced prediction accuracy and an enhanced levelof explainability with respect to the plurality of trained ML models. The model explainability phasemay be performed by a first processing module 611 of the model optimization device 6.

[0120] At a step 401, specific foot data is collected. Such specific foot data may be collected from atleast one sensor 1b disposed at least within a shoe insole of a user. Such user may for examplecorrespond to a user of the diabetes monitoring system 1 as disclosed in figure 1. In particular, suchuser may not be included in the plurality of individuals whose collected data served to determine thetrained ML models during the modeling phase 300. The specific foot data may then be collectedwhen the diabetes monitoring system 1 is used by the user. In particular, such specific foot datacorresponds to unlabeled data.

[0121] In an embodiment, the specific foot data corresponds to foot pressure data collected from atleast one pressure sensor 1b disposed within the shoe insole of the user. In the rest of the presentdisclosure, the specific foot data will refer to the foot pressure data. A person skilled in the art wouldunderstand that such disclosure may apply to other types of specific data, such as foot temperature data or foot humidity data for example.

[0122] At a step 402, a plurality of trained ML models are retrieved. Such plurality of trained MLmodels may be retrieved from the model storage 5. The trained ML models stored in the modelstorage 5 may include trained ML models stored after a modeling phase 300 performed by amodeling device 3 for example. Such model storage 5 may also include any other trained ML modelresulting from any other supervised learning process. In particular, such plurality of trained MLmodels have multi data input. For example, each trained ML model as retrieved may take medical information, foot pressure data, foot temperature data and foot humidity data as data inputs.

[0123] The foot pressure data collected at step 401 from the user enables to run steps 410 to 470.

[0124] At a step 410, a feature importance analysis is performed on the plurality of trained ML models,based on the foot pressure data. In other words, at step 410, it is determined the contribution of eachfeature (that is, input data) on the overall predictive performance of each trained ML model. Forexample, for each of the trained ML model resulting from the modeling phase 300, the featureimportance analysis performed at step 410 enables to determine the relative importance of the foot pressure data as a data input on such trained ML model with respect (or in the vicinity of) the foot temperature data, the foot humidity data and the medical information.

[0125] To that end, step 410 may include performing an ensemble learning process, for exampleusing a decision tree (or random forest) model. In an embodiment, the feature importance analysismay also be performed at step 410 using a Gradient Boosting model, such as XGBoost or LightGBM.

[0126] During step 410, feature importance scores may be determined for each of the plurality oftrained ML models. Such feature importance scores may for example depend on a relative importance of the foot pressure data in the predictive performance of the corresponding trained ML models.

[0127] At a step 420, a transfer learning step is performed, in order to adapt the plurality of trainedML models to work with the specific foot data only. To that end, Generative Adversarial Networks(GANs) may be performed at step 420. In other words, step 420 generates a plurality of generatedML models configured for decision-making based on solely the foot pressure data as data input.

[0128] At a step 430, a hyperparameter tuning step is performed on each of the generated MLmodels. Step 430 includes retuning the hyperparameters of each of the generated ML models at step420 in order to adapt such generated ML model to a context of having only foot pressure data as adata input. To that end, step 430 may use an adaptive optimizer, such as an Adam optimizer. Inanother embodiment, step 430 may also be performed using a Stochastic Gradient Descent (or SGD).

[0129] At a step 440, a cross-validation step is performed. Such step 440 enables to evaluate theperformance of each of the generated ML model after step 430, using only foot pressure data. Tothat end, step 440 may include performing a K-fold cross validation for each of the generated ML model.

[0130] At a step 450, an explainability step is performed. For each of the generated ML model,explainability parameters may be determined. Such explainability parameters may for exampleinclude values of explainability metrics such as Root Mean Square Error (or RMSE), Mean AbsoluteError (or MAE), accuracy and / or an F1-score associated to each of the generated ML model.

[0131] At a step 460, a performance comparison step is performed. Step 460 aims at comparing theprediction performance of each of the generated ML model following step 450 with respect to apredetermined baseline model, in order to ensure that each generated ML model is adding predictionvalue. In other words, the comparison of each of the generated ML model with the baseline modelensures that only generated ML models that add a significant predictive value for decision-making indiabetes diagnosis will be further considered in the rest of the method (and notably for the creationof a meta-model for diabetes diagnosis, which will be described further).

[0132] To that end, a simple baseline model may be predetermined with respect the plurality ofgenerated ML models to be evaluated. Such baseline model may for example have a naive, simple or basic prediction approach of diabetes based on foot pressure data. The performance of eachgenerated ML model may be compared to the performance of such baseline model to ensure theadded prediction value (that is, a significant improvement in prediction performance) of the generatedML models. Such added prediction value may be ensured by evaluating various performance metricssuch as accuracy or error rates for example.

[0133] At such step 460, a meta-model based on the plurality of generated ML models may begenerated. Such meta-model corresponds to a model of the generated ML models (that is, a modelof models). The generated meta-model represents a prediction tool leveraging the best attributes ofeach contributing model for enhanced predictive performance. To that end, the meta model is created based on the comparisons performed between each of the plurality of generated ML models and the baseline model. The generated ML models that are considered to add a significant predictive value based on such comparisons are referred to as specific ML models and are used during step 460 togenerate the meta-model. Such generated meta-model then incorporates insights, strengths and keycharacteristics from the specific ML models. The most performant aspects of each of the specific ML models are combined, resulting in a meta-model being a robust, accurate and explainable prediction tool. In particular, the meta-model has a prediction performance superior to each of the generated ML models (and to each of the specific ML models).

[0134] At a step 470, a model report step is performed. Such model report step 470 may includegathering and reporting data related to each of the specific ML models, associated to input datastructure, model hyperparameters and parameters, performance metrics, explainability metrics etc.as determined at the previous steps 420 to 460 for such specific ML models. The model report step470 may also include data related to the meta-model of the plurality of specific ML models as generated at step 460.

[0135] Such model report step 470 may for result in an XML file corresponding to a model report oneach of the plurality of specific ML models and / or a model report on the meta-model based on suchspecific ML models. In particular, at such step 470, specific ML models are specific to foot pressuredata (or more generally to the specific foot data). The model report generated at step 470 may bestored in a report storage 62. Such report storage 62 may be included in the model optimization device 6 as a storage unit of the model optimization device 6. In an embodiment, such report storage 62 may also be a distant storage unit with respect to the model optimization device 6.

[0136] The model report generated at step 470 may also include parameters and / orhyperparameters related to algorithms used in the processing steps 410 to 470 performed during themodel optimization phase 400.

[0137] For example, if Random Forests algorithms are used during the feature importance analysisat step 410, the model report determined at step 470 may include one or several elements of thefollowing data related to such Random Forests algorithms:- Number of Trees,- Maximum Depth,- Minimum Sample Split,- Minimum Sample Leaf, and- Features to Consider.

[0138] For example, if Generative Adversarial Networks (GAN) models are used during the transferlearning at step 420, the model report determined at step 470 may include one or several elementsof the following data related to such Generative Adversarial Networks (GAN):- Latent Space Dimension,- Generator Learning Rate,- Discriminator Learning Rate,- Batch Size,- Epochs,- Beta1, and- Beta2.

[0139] For example, the model report determined at step 470 may include one or several elementsof the following data related to the hyperparameter tuning at step 430:- Learning Rate,- Regularization Terms,- Optimizer,- Batch Size, and- Epochs.

[0140] For example, the model report determined at step 470 may include one or several of theexplainability parameters determined at step 450.

[0141] For example, the model report determined at step 470 may also include one or severalelements among the Validation Split, the Cross Validation method used at step 440, the scoring function, the early stopping criteria.

[0142] For example, the model report determined at step 470 may include the specific foot data (e.g.,the foot pressure data) and the feature names, as represented in [Table 6] of the Examples section.

[0143] More generally, the model report generated at step 470 includes all parameters, data,hyperparameters, specific ML models that have a significant role in the shaping of the meta-model.

[0144] At an optional step 480, a continuous learning in the model explainability phase may beperformed, so as to improve and update the generated ML models (and notably the specific MLmodels) and the corresponding meta-model (and the resulting model report of step 470) as updatedspecific foot data is collected at step 401. For example, at each occurrence of step 401 of specificfoot data collection (for example, when the user uses the diabetes monitoring system 1 and runs the system 1 to obtain a diabetes monitoring result), an occurrence of the model explainability phase may be performed as an unsupervised learning process to output a plurality of specific ML modelsspecific to such collected specific foot data. For example, each time the user wears the foot pressuresensors and foot pressure data is collected at step 401, the model explainability phase 400 runs an unsupervised learning stage, generates a meta-model associated to the plurality of specific (andpressure-specific) ML models, such meta-model being configured to specify a predicted outputrelated to diabetes monitoring and / or diagnosis, for example, “Onset of Diabetics” or “Non-Diabetics”, based on the inputted foot pressure data.

[0145] It is now referred to figure 5. Figure 5 illustrates steps of a model reliability assessment phase500, for determining an optimized ML model for decision-making in diabetes diagnosis. In particular,such model reliability assessment phase 500 aims at assessing a prediction reliability level of eachof the plurality of specific ML models as determined during the model explainability phase 400.Similarly, such model reliability assessment phase 500 may aim at assessing a prediction reliabilitylevel of the generated meta-model resulting from such plurality of specific ML models. The furtherdescription of the model reliability assessment phase 500 will be illustrated with the specific MLmodels. Equivalently, such phase 500 applies to the generated meta-model resulting from suchplurality of specific ML models. Such model reliability assessment phase 500 may be performed bythe model optimization device 6, and more particularly by a second processing module 612 of the model optimization device 6.

[0146] In particular, the model reliability assessment phase 500 may rely on a hybrid framework,including both the use of neutrosophic sets and intuitionistic fuzzy logic.

[0147] At a step 510, data related to the plurality of specific ML models and / or data related to themeta-model are retrieved. In particular, the model report data generated at step 470 is retrieved fromthe report storage 62, for further processing by the second processing module 612 during the modelreliability assessment phase 500. The prediction output of each of the specific ML models and / or ofthe meta-model at this stage may be in a binary form, for example as “Diabetes” and “Non-Diabetes”outputs forming the binary output set of each specific ML model and / or of the meta-model. Each ofthe specific ML models and / or the meta model also includes a plurality of outputs representing the parameters, hyperparameters, variables and / or data input associated to such ML model. More generally, such outputs associated to each specific ML model is contained in the model report data retrieved from the report storage 62. Such outputs belong to a universe of discourse X associated to the concerned ML model.

[0148] At a step 520, neutrosophic sets are integrated to each of the specific ML models (orequivalently, to the meta-model). In other words, at step 520, each of the specific ML models aremodified to represent uncertainty, indeterminacy and falsity levels in their model outputs. Forexample, for each specific ML model as described in the model report retrieved at step 510, an output layer of the specific ML model may be modified to integrate neutrosophic sets.

[0149] For each specific ML model, the neutrosophic set may for example be defined as:

[0150] NS = {^^, ^^^(^), ^^^(^), ^^^(^)^ ∣ ^ ∈ ^}

[0151] where:- NS is the neutrosophic set,- X is the universe of discourse of the specific ML model,- x is a possible output of the universe of discourse X,- TNS(x), INS(x) and FNS(x) are respectively the degrees of truth, indeterminacy and falsity ofthe output x.

[0152] In particular, the values TNS(x), INS(x) and FNS(x) may not bounded by their sum (e.g., suchvalues are not necessarily probabilistic values which sums to 1). For a given ML model (such thespecific ML models and the resulting meta-model), the universe of discourse X may be understood as the entire set of considered possible elements. For example, for a given ML model, the universe of discourse X may refer to the entire space of possible feature vectors and engaged hyperparameters. An output x of a specific ML model may for example be a parameter, a hyperparameter or a feature vector for example.

[0153] At step 520, such neutrosophic sets NS are integrated to each of the specific ML models.

[0154] At a step 530, intuitionistic fuzzy sets are integrated to each of the specific ML models. Inother words, at step 530, each of the specific ML models are modified to represent degrees ofmemberships and non-membership in their model outputs. Compared to basic fuzzy sets, theintuitionistic fuzzy sets introduces a degree of hesitation or non-membership in each specific MLmodel as an independent measurement, distinct from the inverse of membership. For example, foreach specific ML model as described in the model report retrieved at step 510, an output layer of thespecific ML model may be modified to integrate intuitionistic fuzzy sets.

[0155] For each specific ML model, an intuitionistic fuzzy set may for example be defined as:

[0156] IFS = {^^, ^^^^(^), ^^^^ (^), ^^^^(^)^ ∣ ^ ∈ ^}

[0157] where:- IFS is the intuitionistic fuzzy set,- X is the universe of discourse of the specific ML model,- x is a possible output of the universe of discourse X,- µIFS(x), ^ IFS(x) and ^^^^(^) are respectively the degrees of membership, non-membershipand hesitation (or equivalently, of certainty) of the output x, with, for all x belonging to X, 0 ≤^^(^) + ^^(^) ≤ 1 and ^^^^(^) = 1 − ^^^^(^) − ^^^^(^).

[0158] At a step 540, the plurality of specific ML models may correspond to a plurality of hybrid MLmodels, which correspond to specific ML models which outputs integrate both neutrosophic sets NSand intuitionistic fuzzy sets IFS. Such hybrid ML model is still specific to specific foot data. At suchstep 540, each hybrid ML model may be defined as the following hybrid set:

[0159] H = {^^, ^^(^), ^^ (^), ^^(^), ^^(^), ^^(^), ^^(^)^ ∣ ^ ∈ ^}

[0160] where:- H is the hybrid set associated to the hybrid ML model,- X is the universe of discourse of the hybrid ML model,- x is a possible output of the universe of discourse X,- µH(x), ^H(x) and ^^(^) are respectively the degrees of membership, non-membership andhesitation of the output x for the hybrid ML model from intuitionistic fuzzy sets,- TH(x), IH(x) and FH(x) are respectively the degrees of truth, indeterminacy and falsity of theoutput x for the hybrid model ML model from neutrosophic sets.

[0161] Each of the plurality of hybrid ML models may then be associated to a hybrid set H.

[0162] In particular, evaluation, explainability, performance metrics and / or any other metricsassociated to the plurality of specific ML models and / or the generated meta-model based on suchplurality of specific ML models may be determined in the context of such hybrid ML models. To thatend, operations on and between the respective hybrid sets H (associated to the plurality of hybridML models) may be defined at step 540. Such operations may for example include union, intersectionand / or complement in or between hybrid sets.

[0163] For example, considering the generated meta-model being a Random Forest-derivated model,an output x1 may correspond to the number of trees of the meta-model and another output x2 may correspond to the foot pressure data in the left foot of the user. The corresponding hybrid set H and its contained values may then define the hesitation levels when the meta-model uses outputs x1 and x2.

[0164] For example, it may be considered two hybrid sets H1 and H2 respectively associated to anytwo hybrid ML models.

[0165] The union set between hybrid sets H1 and H2 may be defined as, for any output x of theuniverse of discourse X of any of the hybrid ML models:

[0166] ^^ ⋃ ^^ = {^^, ^^^ ⋃ ^^ (^), ^^^ ⋃ ^^ (^), ^^^ ⋃ ^^ (^), ^^^ ⋃ ^^ (^), ^^^ ⋃ ^^(^), ^^^ ⋃ ^^ (^)^ ∣ ^ ∈ ^}

[0167] with:- ^^^ ⋃ ^^(^) = max(^^^(^), ^^^(^))- ^^^ ⋃ ^^ (^) = min(^^^(^), ^^^(^))-f and g are predetermined functions.

[0168] The intersection set between hybrid sets H1 and H2 may be defined as, for any output x of theuniverse of discourse X of any of the hybrid ML models:

[0170] with:- ^^^ ⋂ ^^(^) = min(^^^(^), ^^^ (^))- ^^^ ⋂ ^^ (^) = max(^^^ (^), ^^^(^))- ^^^ ⋂ ^^ (^) = ℎ (^^^(^), ^^^(^))- ^^^ ⋂ ^^ (^) = min(^^^(^), ^^^(^))--^^^ ⋂ ^^(^) = max(^^^(^), ^^^(^))- h and i are predetermined functions.

[0171] The complement set of a hybrid set H may be defined as, for any output x of the universe ofdiscourse X of the hybrid ML model:

[0173] At a step 540, it is thus possible to define, for each output x of each specific ML model (orequivalently, of the meta-model), membership values based on each hybrid ML model associated tosuch specific ML model. In particular, such membership values associated to an output x maycorrespond to numerical values in the hybrid set H associated to such hybrid ML model, that is to, ^^(^), ^^(^), ^^(^), ^^ (^), ^^(^), ^^(^).

[0174] Thus, at step 540, such membership values may be included between 0 and 1 and may reflectvarious degrees of certainty, indeterminacy, hesitation etc. with respect to each output x of thespecific ML model (or of the meta-model).

[0175] For example, it is considered a given specific ML model corresponding to a specific randomforest model, as retrieved at step 510. A considered output x of such retrieved specific random forestmodel may correspond to a hyperparameter being the number of trees. For such output x, forexample, ^ may take values between 0 and 1 and may reflect the degree of certainty about thesetting value of x, ^^(^) = 1 meaning that the setting is certain for example.

[0176] In another example, a considered output x of such retrieved specific ML model maycorrespond to a parameter being the learning rate. For such output x, for example, ^ may take valuesstrictly between 0 and 1 which means that the optimal learning rate is uncertain, ^ may take valuesbetween 0 and 1, ^^(^) = 1 meaning that the specific ML model has a bad learning rate.

[0177] In particular, the hybrid ML models and the membership values depend on the specific MLmodels and the model construction of such specific ML models as retrieved. Thus, updated footspecific data potentially leads to updated specific ML models (for example obtained via continuouslearning at step 480) and thus leads to updated hybrid ML models and membership values for eachmodel output.

[0178] At a step 550, assessment thresholds are defined with respect to the membership values, foreach specific ML model (or equivalently, for the meta-model). Such step 550 notably relies on thecorresponding hybrid ML models and the membership values as determined at step 540.

[0179] In particular, assessment thresholds may be preset and fixed values comprised between 0and 1. Such assessment thresholds enable to assess (or evaluate) the reliability of each specific MLmodel (or of the meta-model), by evaluating the membership values associated to such specific MLmodel outputs.

[0180] In an embodiment, assessment thresholds for each model output may include at least a highassessment threshold and a low assessment threshold. For example, a high assessment thresholdfor a given output may be set at 0,8 and a low assessment threshold may be set at 0,3. In anembodiment, assessment thresholds may be set at different values for different specific ML models and / or for different model outputs.

[0181] Such assessment thresholds may notably be reset for each retrieved plurality of specific MLmodels (that is, for each new iteration of step 510).

[0182] At a step 560, each of the plurality of specific ML models (or equivalently, the meta-model) isevaluated based on the assessment thresholds.

[0183] To that end, at step 560, an output of the considered specific ML model is compared with theassessment thresholds and a reliability level is associated (or matched) to such output (or to thespecific ML model) depending on such comparison. In particular, the reliability level may take one value among a high reliability level, a medium reliability level and a low reliability level.

[0184] In particular, the reliability level may be associated to each output based on each comparison.The reliability level may for example be associated to each output of the assessed meta-model. In another embodiment, the reliability level may be associated to the specific ML model based on the comparison result of one output, all outputs or a minimal number of outputs. For example, if more than half the number of outputs of a specific ML model have a comparison resulting in a high reliability level associated to such outputs, a high reliability level may be associated to the specific ML model. To simplify the further description, it is considered that the reliability level is associated to the specificML model (in other words, it is considered that the result of the comparison of one output with theassessment thresholds results in evaluating the specific ML model).

[0185] At step 560, if the output exceeds the high assessment threshold (in other words, if all or aminimal number of membership values exceed the high assessment threshold), the specific machine learning model is evaluated as having a high reliability level.

[0186] At step 560, if the output is below the low assessment threshold (in other words, if all or amaximal tolerated number of membership values are below the low assessment threshold), thespecific machine learning model is evaluated as having a low reliability level.

[0187] At step 560, if the output is between the high assessment threshold and the low assessmentthreshold (in other words, if all or a predefined number of membership values are between the highassessment threshold and the low assessment threshold), the specific machine learning model isevaluated as having a medium reliability level.

[0188] At such stage of step 560, each of the plurality of specific ML models is associated to areliability level among a high reliability level, a medium reliability level and a low reliability level.

[0189] At step 560, if none of the specific ML models is associated to a medium reliability level, step570 is directly performed and will be further described. If at least one specific ML model has a mediumreliability level, a step 561 is further performed.

[0190] At a step 561, specific ML models having a medium reliability level are transmitted to acomplementary model assessment unit. In another embodiment, outputs having a medium reliability level are transmitted to the complementary model assessment unit. Such complementary modelassessment unit may for example correspond to an expert terminal 7, as represented on figure 1. Inparticular, such complementary model assessment unit may be configured to evaluate thetransmitted specific ML models based on at least one medical assessment threshold. In order toevaluate the transmitted specific ML models (or the transmitted outputs), the complementary model assessment unit may further be provided with diabetes analysis and / or complementary informationrelated to such outputs and / or such specific ML models. Such medical assessment threshold mayfor example result from medical knowledge, medical information and / or any other medical source ofexpertise enabling to assess the diagnosis result reliability of the specific ML models. Such medicalassessment threshold may for example be stored in the memory unit 70 of the expert terminal 7.

[0191] At step 561, a complementary model assessment is performed by the complementary modelassessment unit on each of the transmitted specific ML models based on the medical assessmentthreshold. If the output of the transmitted specific ML model exceeds the medical assessmentthreshold, such transmitted specific ML model is evaluated as having a high reliability level. In theoutput of the transmitted specific ML model is below the medical assessment threshold, suchtransmitted specific ML model is evaluated as having a low reliability level. Such evaluation may beperformed either by the complementary model assessment unit or by the model optimization device 6.

[0192] At a step 570, following step 560 and possibly step 561, each of the plurality of specific MLmodels is evaluated as having either a high reliability level or a low reliability level.

[0193] At such step 570, the specific ML models evaluated as having a low reliability level are notfurther exploited by the model optimization device 6. In an embodiment, such specific ML modelshaving a low reliability level may be transmitted to a model storage for further processing. Forexample, such specific ML models may be transmitted to the model management device 4 asrepresented on figure 1, for further storage in the model storage 5 and for further use in futureiterations of the model explainability phase.

[0194] At such step 570, the specific ML models evaluated as having a high reliability level areconsidered by the model optimization device 6 as a candidate specific ML model. In another embodiment, the outputs evaluated as having a high reliability level are considered by the model optimization device 6 for determining the optimized ML model.

[0195] At a step 580, the optimized ML model is selected among all specific ML models evaluatedas having a high reliability level. For example, such optimized ML model may correspond to thespecific ML model having the highest average membership values. The optimized ML model maycorrespond to the specific ML having the highest prediction accuracy or the highest performancemetrics overall. The optimized ML model may also be determined as a meta-model of the specificML models evaluated as having a high reliability level, using ensemble learning for example. Theoptimized ML model may also be determined as the meta-model of the specific ML models if suchmeta-model is evaluated as having a high reliability level. The optimized ML model may also bedetermined as the ML model build based on the outputs having a high reliability level.

[0196] Thus, at step 580, an optimized ML model is determined for decision-making in diabetesdiagnosis or monitoring. In particular, such optimized ML model is determined as having satisfactory explainability level, accuracy level and reliability level based on specific foot data as data input.

[0197] Thus, such optimized ML model as determined may be exploited, as described at step 500aof figure 2, to provide a diabetes monitoring result to the user. In particular, such diabetes monitoringresult is specific and optimized to the specific foot data (e.g., the foot pressure data) collected fromsuch user and processed by an unsupervised means in a model explainability phase 400 and in areliability assessment phase 500. Such optimized ML model thus enables to provide more accurate,reliable and explainable results to the user regarding its diabetes diagnosis or monitoring, suchresults taking the specific foot data of the user into account. Examples

[0198] [Table 1] illustrates an example of structured data, corresponding to collected data during themodeling phase 300, at steps 302 and 303.

[0199] [Table 1]Data Source Parameter Data Type Range / Unit DescriptionDomain Expert GlycemicNumerical mg / dL Predefined bloodKnowledge Thresholds glucose levels indicating diabetes risk Diabetes LevelRisk Categories Categorical Low / Medium / High Classification ofTaxonomy diabetes risk levels Humidity RelativeNumerical % Humidity level detectedSensor Data Humidity in the foot area Pressure PlantarNumerical kPa Pressure exerted on aSensor Data Pressure plantar area of the foot TemperatureTemperature Numerical ° C Temperature detectedSensor Data in the foot area

[0200] [Table 2] and [Table 3] illustrate examples of structured data, corresponding to collectedmedical information data collected during the modeling phase 300, at step 303.

[0201] [Table 2]Data Source Parameter Data Type Range / Unit DescriptionDomain ExpertGlycemic Control Numerical mg / dL orEstablished blood Knowledge Thresholds mmol / L glucose levels critical for diabetes management Domain Expert DiagnosisCategorical Binary / Scale Criteria for diabetesKnowledge Criteria diagnosis such as fasting glucose levels,HbA1c percentages. Domain Expert Risk StratificationNumerical Score / Level Metrics for assessingKnowledge Metrics diabetes complication risks Domain Expert TreatmentTextual N / A Guidelines on managingKnowledge Protocols various stages of diabetesDomain ExpertSymptomatology Categorical List ofCommon symptoms Knowledge Symptoms associated with diabetes onset or progression

[0202] [Table 3]Data Source Parameter Data Type Range / Unit DescriptionDiabetes LevelDiabetes Type Categorical Type 1 / Type 2Classification of Taxonomy etc. diabetes Diabetes Level Blood GlucoseNumerical Mg / dL or mmol / L Defined ranges forTaxonomy Levels fasting, postprandial and random glucose tests Diabetes Level HbA1cNumerical % Average blood glucoseTaxonomy Percentage over the past three months Diabetes Level InsulinNumerical Unitless Quantitative measure ofTaxonomy Sensitivity insulin efficiency Diabetes Level ComplicationCategorical Low / Medium / High Risk levels for potentialTaxonomy Risk Levels diabetes-related complications Diabetes Level TreatmentCategorical Low / Medium / High Effectiveness ofTaxonomy Efficacy Levels different treatment regimens

[0203] [Table 4], [Table 5] and [Table 6] illustrate examples of structured data, corresponding tocollected sensor-based data collected during the modeling phase 300, at step 302. In particular,[Table 6] illustrates an example of specific foot data corresponding to foot pressure data collectedduring the model optimization phase 400 at step 401.

[0204] [Table 4]Data Source Parameter Data Type Range / Unit DescriptionFoot Humidity AmbientNumerical % Measurement of the moistureSensor Humidity level in the air surrounding the foot Foot HumidityInsole Humidity Numerical % Measurement of the moistureSensor level within the shoe in the foot area Foot Humidity TemporalNumerical % over time Tracks changes in humiditySensor Variations over a predefined time period Foot Humidity HumidityNumerical % Detection fluctuations andSensor Fluctuations their speed in humidity, which could reflect sweating or other conditions

[0205] [Table 5]Data Source Parameter Data Type Range / Unit DescriptionFoot Foot SurfaceNumerical °C or °F Measurement of theTemperature Temperature temperature at the surface of Sensor the foot Foot AmbientNumerical °C or °F Measurement of theTemperature Temperature environmental temperature in Sensor the foot area Foot TemporalNumerical °C or °FTracks changes in Temperature Temperature over timetemperature over a predefinedSensor Trends time period Foot TemperatureNumerical °C or °F Detection fluctuations inTemperature Variability temperature, which could Sensor reflect health issues

[0206] [Table 6]Data Source Parameter Data Type Range / Unit DescriptionFoot Pressure PlantarNumerical kPa or psi Pressure exerted on differentSensor Pressure areas of the foot Foot Pressure PressureNumerical DistributionVisualization of pressure Sensor Distribution Map points across the foot Foot Pressure PressureNumerical kPa or psi Changes in pressure duringSensor Variability different activities Foot Pressure DynamicNumerical kPa or psi Patterns of pressure changeSensor Pressure during movement Patterns Foot Pressure Peak PressureNumerical kPa or psi Highest pressure areasSensor Points identified by sensors Reference Signs List

[0207] 1: diabetes monitoring system- 1a: individual sensor- 1b: user sensor- 2: medical knowledge unit- 3: modeling device- 4: model management device- 5: model storage- 6: model optimization device- 7: expert terminal- 60, 70: memory units- 31, 41, 61, 71: processing units- 31a, 31b, 31c, 31d, 611, 612: processing modules- 62: report storage- 8: user terminal

Claims

Claims

1. Method (200) performed by a computer system for monitoring diabetes, the method (200)comprising:a model explainability phase (400), comprising steps of:- collecting (401) specific foot data from at least one sensor (1b) disposed at least within ashoe insole of a user,- collecting (402) a plurality of trained machine learning models configured for decision-makingin diabetes diagnosis based on a plurality of collected data,- based on at least the plurality of trained machine learning models, generating a plurality ofspecific machine learning models, said specific machine learning models being specified fordecision-making in diabetes diagnosis based on the specific foot data,and a model reliability assessment phase (500), comprising steps of:- evaluating (560) a reliability of each of the specific machine learning model based onneutrosophic sets (NS) and intuitionistic fuzzy sets (IFS),- based on the result of the evaluation step (560), selecting (580) the optimized machinelearning model for monitoring diabetes based on the specific foot data.

2. Method (200) according to claim 1, wherein the specific foot data corresponds to footpressure data collected from at least one pressure sensor (1b) disposed within the shoe insole of theuser.

3. Method (200) according to any one of the preceding claims further comprises:a modeling phase (300), comprising steps of:- collecting data including sensor-based data (302) collected from a plurality of sensors (1a)disposed at least within a shoe insole of a plurality of individuals, and medical information(303) related to diabetes diagnosis, -based on the collected data, determining the plurality of trained machine learning models.

4. Method (200) according to claim 3, wherein determining the plurality of trained machinelearning models further comprises a step of: preprocessing (310) the collected data, said preprocessing including steps of cleaning,normalizing, feature extracting and / or feature selecting.

5. Method (200) according to any one of claims 3 and 4, wherein determining the plurality oftrained machine learning models further comprises steps of: selecting (301) a plurality of initial machine learning models, andfor each selected initial machine learning model:splitting (320) the collected data into at least a training dataset and a test dataset,training (330) the selected initial machine learning model using the training dataset,evaluating (340) a performance of the selected initial machine learning model after thetraining step (330), using the test dataset,depending on the performance evaluation step (340), adjusting (350) parameters and / orhyperparameters of the selected initial machine learning model, the trained machine learning modelresulting from said adjustment step (350).

6. Method (200) according to any one of claims 3 to 5, wherein the modeling phase (300)further comprises: collecting updated data including updated sensor-based data (302) and / or updated medicalinformation (303),performing (360) a continuous learning process based on said collected updated data.

7. Method (200) according to any one of the preceding claims, wherein the modeling phase(300) is based on supervised learning processes using at least the collected data.

8. Method (200) according to any one of the preceding claims, wherein the collected dataincludes sensor-based data (302), said sensor-based data (302) including at least one elementamong: -foot humidity data collected on at least one humidity sensor (1a) disposed within the shoeinsole of a plurality of individuals, -foot temperature data collected on at least one temperature sensor (1a) disposed within theshoe insole of the plurality of individuals, and- foot pressure data collected on at least one pressure sensor (1a) disposed within the shoeinsole of the plurality of individuals.

9. Method (200) according to claim 8, wherein the foot humidity data includes at least oneelement among: -ambient humidity data, related to a moisture level of the air surrounding the humidity sensor(1a), -insole humidity data, related to a moisture level within the shoe insole of each of the pluralityof individuals, -humidity variation data, related to variations of a humidity level over a predefined period, and- humidity fluctuation data, related to paces of changes of the humidity level.

10. Method (200) according to claim 8, wherein the foot temperature data includes at leastone element among: -ambient temperature data, related to a temperature of the air surrounding the temperaturesensor (1a),- foot surface temperature data, related to a temperature at a surface of a foot of each of theplurality of individuals, -temperature variation data, related to variations of a temperature over a predefined period,and -temperature fluctuation data, related to paces of changes of temperature.

11. Method (200) according to any one of claims 2 and 8, wherein the foot pressure dataincludes at least one element among: -plantar pressure data, related to a pressure level exerted on different plantar regions of afoot of one among the user and the plurality of individuals,- pressure distribution data, related to a mapping of pressure points on the foot of one amongthe user and the plurality of individuals,- pressure variation data, related to variations of the pressure level over a predefined period,- pressure variation pattern data, related to patterns of pressure changes over predefinedtimes, and- peak pressure data, related to extremal pressure values exerted in the different plantarregions of the foot of one among the user and the plurality of individuals.

12. Method (200) according to claim 3, wherein the medical information (303) related todiabetes diagnosis includes at least one element among:- predefined glycemic level thresholds,- predefined diabetes diagnosis criteria,- predefined metrics for assessing diabetes complication risks,- predefined diabetes treatment protocols for different diabetes stages,- predefined lists of symptoms associated to different diabetes stages,- lists of diabetes classification types,- values of blood glucose test levels,- glycated hemoglobin – HbA1c – average values,- insulin efficiency values,- risk level values for diabetes complication risks,- effectiveness values of treatment protocols.

13. Method (200) according to any one of the preceding claims, wherein the modelexplainability phase (400) is based on unsupervised learning processes.

14. Method (200) according to any one of the preceding claims, wherein the modelexplainability phase (400) is performed using at least one element among Generative AdversarialNetworks – GAN -, random forest algorithms, pattern recognition algorithms.

15. Method (200) according to any one of the preceding claims, wherein the modelexplainability phase (400) further includes a step of:- performing (410) a feature importance analysis on the plurality of trained machine learningmodels based on the specific foot data.

16. Method (200) according to any one of the preceding claims, wherein the modelexplainability phase (400) further includes a step of: -performing (420) a transfer learning on the plurality of trained machine learning models, inorder to adapt the plurality of trained machine learning models for decision-making in diabetes diagnosis based on the specific foot data as an only data input.

17. Method (200) according to any one of the preceding claims, wherein the modelexplainability phase (400) further includes a step of: -performing (430) a hyperparameter tuning of each of the plurality of trained machine learningmodels, in order to adapt a learning process of each of the plurality of trained machine learning model to learn with the specific foot data as an only data input.

18. Method (200) according to any one of the preceding claims, wherein the modelexplainability phase (400) further includes a step of: -performing (440) a cross-validation to evaluate a model performance of each of the pluralityof trained machine learning models, using the specific foot data as an only data input.

19. Method (200) according to any one of the preceding claims, wherein the modelexplainability phase (400) further includes a step of: -determining (450) explainability parameters related to an explainability of each of the pluralityof trained machine learning models.

20. Method (200) according to the precedent claim, wherein the explainability parametersinclude at least one element among: a Root Mean Square Error – RMSE- value, a Mean AbsoluteError – MAE- value, an accuracy value, an F1-score value.

21. Method (200) according to any one of the preceding claims, wherein the modelexplainability phase (400) further includes a step of: -determining (460) a meta-model of the plurality of specific machine learning models basedon a comparison of a performance of each of the plurality of specific machine learning models with a predefined baseline model.

22. Method (200) according to any one of the preceding claims, wherein the modelexplainability phase (400) further includes a step of:- generating (470) model report data related to the plurality of specific machine learningmodels and / or to a meta-model generated based on said plurality of specific machinelearning models.

23. Method (200) according to any one of the preceding claims, wherein the modelexplainability phase (400) further includes steps of:- collecting (401) updated specific foot data,- performing (480) a continuous learning based on said collected updated specific foot data.

24. Method (200) according to any one of the preceding claims, wherein the model reliabilityassessment phase (500) further comprises steps of:- integrating (520) neutrosophic sets to the specific machine learning models,- integrating (530) intuitionistic fuzzy sets to the specific machine learning models,- based on the integrated neutrosophic sets and intuitionistic fuzzy sets, determining (540)membership values associated to outputs of each of the specific machine learning models,- establishing (550) assessment thresholds with respect to the membership values, saidassessment thresholds including at least a high assessment threshold and a low assessmentthreshold.

25. Method (200) according to claim 24, wherein the model reliability assessment phase(500) further comprises steps of: for each specific machine learning model, -evaluating (560) the specific machine learning model based on the assessment thresholds,- if a membership value of an output of the specific machine learning model exceeds the highassessment threshold, evaluating the specific machine learning model as having a highreliability level, -if the membership value of the output of the specific machine learning model is below thelow assessment threshold, evaluating the specific machine learning model as having a lowreliability level, -if the membership value of the output of the specific machine learning model is between thehigh assessment threshold and the low assessment threshold, evaluating the specificmachine learning model as having a medium reliability level.

26. Method (200) according to claim 25, wherein the model reliability assessment phase(500) further comprises steps of: -transmitting (561) the specific machine learning models having a medium reliability level toa complementary model assessment unit,the complementary model assessment unit being configured to evaluate the transmittedspecific machine learning models based on a medical assessment threshold,- if the output of the transmitted specific machine learning model is below the medicalassessment threshold, transferring said transmitted specific machine learning model to a model storage storing the plurality of trained machine learning models,- if the output of the transmitted specific machine learning model is above the medicalassessment threshold, evaluating the specific machine learning model as having a highreliability level.

27. Method (200) according to any one of claims 25 and 26, wherein the optimized machinelearning model is selected (580) among the specific machine learning models having a high reliabilitylevel.

28. Method according to any one of the preceding claims further comprises a diabetesmonitoring phase (500a) including a step of:- transmitting a diabetes monitoring result to a final user interface, said diabetes monitoringresult being determined based on an output of the optimized machine learning model basedon the specific foot data.

29. Method (200) according to claim 28, wherein the final user interface may be at least oneelement among: -a final user wearable,- a final user terminal (8),- a human-machine interface,- a cloud interface.

30. Diabetes monitoring system (1) comprising computer means for performing a method(200) according to one of claims 1 to 29.

31. Computer software comprising instructions to implement at least a part of a method (200)according to one of claims 1 to 29 when the software is executed by a processor (61).

32. Computer-readable non-transient recording medium on which a software is registered toimplement a method (200) according to one of claims 1 to 29 when the software is executed by aprocessor (61).

Images