Method for determining the physiological age of a subject

Abstract

A computer-implemented method for determining a physiological age of a subject is disclosed, the method comprising applying a trained model on a set of values ​​including at least values ​​of a biological variable for the subject, the trained model being configured to predict the chronological age of the subject based on these sets of values ​​to obtain a predicted age of the subject, wherein a physiological age corresponds to the predicted age.

Description

[Technical field]

[0001] Technical Field The present disclosure relates to a computer-implemented method for determining the physiological age of a subject and detecting premature aging in said subject. [Background technology]

[0002] 2. Background of the Invention The aging of the population, with its socio-economic impacts, is becoming a major issue for all societies. For many years, the main goal of research on aging was to extend lifespan, but current research is focused on healthy aging and the need to consider the organism as a whole to maintain essential capacities and vital functions. This shift has been accompanied by the emergence of the geroscience paradigm, which considers aging to be the major risk factor common to all chronic diseases and that age-related molecular dysfunction is responsible for functional decline (Kennedy, BK et al., "Geroscience: Linking Aging to Chronic Disease." Cell 159, 709-713 (2014)). However, whereas the original perspective is quite integral, the majority of geroscience research is carried out at the cellular or even molecular scale.

[0003] At the organism level, aging is the result of multifactorial process dysfunctions that cannot be described by a single variable and that involve a multitude of mechanisms, most often highly interdependent. Their accumulation on the health trajectory can vary greatly between people of the same chronological age (Li, Q. et al., "Homeostatic dysregulation proceeds in parallel in multiple physiological systems." Aging Cell 14, 1103-1112 (2015)). Therefore, early monitoring of physiological age at the individual level is crucial to capture these dysfunctions, which usually appear even before subtle clinical symptoms appear. The aim is not only to provide the most appropriate individualized recommendations and interventions to achieve healthy aging, but also to evaluate the effectiveness of anti-aging therapies.

[0004] Recently, deep learning approaches have been used to predict chronological age, but the models have limitations in their explainability (Putin, E. et al., "Deep biomarkers of human aging: Application of deep neural networks to biomarker development." Aging 8, 1021-1033 (2016); Cohen, AA, Morissette-Thomas, V., Ferrucci, L., and Fried, LP, "Deep biomarkers of aging are population-dependent." Aging 8, 2253-2255 (2016)). However, explainability is an essential requirement for both acceptability and applicability for medical applications, as well as for generating physiopathological hypotheses. [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] Kennedy, BK et al., "Geroscience: Linking Aging to Chronic Disease." Cell 159, 709-713 (2014) [Non-Patent Document 2] Li, Q. et al., “Homeostatic dysregulation proceedings in parallel in multiple physiological systems.” Aging Cell 14, pp. 1103–1112 (2015) [Non-Patent Document 3] Putin, E. et al., “Deep biomarkers of human aging: Application of deep neural networks to biomarker development.” Aging 8, pp. 1021–1033 (2016) [Non-Patent Document 4] Cohen, AA, Morissette-Thomas, V., Ferrucci, L. and Fried, LP, "Deep biomarkers of aging are population-dependent." Aging 8, pp. 2253-2255 (2016) [Non-Patent Document 5] Lundberg, Scott et al., "Consistent individualized feature attribution for tree ensembles," 2019 Summary of the Invention [Problem to be solved by the invention]

[0006] SUMMARY OF THE PRESENT APPLICATION In this context, it is an object of the present disclosure to provide improved methods for assessing the physiological age of a subject and for detecting premature aging.

[0007] Another objective of this disclosure is to propose an explainable machine learning framework. [Means for solving the problem]

[0008] Accordingly, a computer-implemented method for determining a physiological age of a subject is disclosed, the method comprising applying a trained model on a set of values ​​including at least values ​​of a biological variable for the subject, the trained model being configured to predict the chronological age of the subject based on these sets of values ​​to obtain a predicted age of the subject, wherein the physiological age corresponds to said predicted age.

[0009] In an embodiment, the method further comprises comparing the subject's physiological age to the subject's chronological age, wherein a positive difference between the subject's physiological age and the chronological age is indicative of premature aging of the subject.

[0010] In an embodiment, the method further comprises comparing the subject's physiological age to a reference age corresponding to the average age predicted by the trained model on a population of the same chronological age as the subject, wherein a positive difference between the subject's physiological age and the reference age indicates premature aging of the subject.

[0011] In an embodiment, the method further comprises comparing the physiological age of the subject with a reference age corresponding to the average age predicted by the trained model on a reference population, and identifying the biological variables that contribute most to the difference when the physiological age of the subject differs from the reference age.The reference population is a group of individuals who have the same chronological age as the individual.The reference population may be a group of individuals that are within a chronological age range of at least 50 years.

[0012] In an embodiment, identifying the biological variables that contribute most to the difference includes determining a SHAP value associated with each of the biological variables and identifying the SHAP value having the highest absolute value.

[0013] In an embodiment, the method further comprises comparing the value of at least one of the biological variables that contributes most to the difference with a reference value of said biological variable for the same chronological age.

[0014] The reference value of a biological variable for a given chronological age may be determined as the average value of the biological variable among a plurality of individuals of said given chronological age, where said biological variable does not contribute to the difference between predicted age and chronological age. The reference value of a biological variable for a given chronological age may also be determined as the average value of the biological variable among a plurality of individuals of said chronological age, where the predicted age is less than or equal to said chronological age. In an embodiment, the method further comprises determining an aging profile for the subject from among a plurality of pre-established aging profiles based on the identified biological or physiological value that contributes most to the difference.

[0015] In an embodiment, the plurality of pre-established aging profiles comprises: - predicting chronological ages of a plurality of individuals of a population by implementing the trained model, the population including a plurality of individuals for each of a plurality of chronological ages; - determining at least the mean expected age of the population; - determining, for a plurality of individuals of the population, the SHAP values ​​of the biological variables that contribute most to the difference between the individual's predicted age and the mean predicted age; - performing clustering on the SHAP values ​​to obtain a finite number of clusters, each cluster corresponding to an ageing profile.

[0016] In an embodiment, the method comprises determining an average predicted age for each of a plurality of chronological ages of a population and determining a SHAP value of the biological variable that contributes most to the difference between the predicted age of the individual and the average predicted age determined for the chronological ages of the individuals, and clustering is performed on said SHAP values.

[0017] In an embodiment, the trained model is an XGboost model that uses a custom loss function that is a function of chronological age.

[0018] In an embodiment, the biological variables include at least some of the following variables: - glycated hemoglobin, - Creatinine in urine, - Cholesterol, - Alanine transaminase (ALT), - mean corpuscular volume, - Aspartate transferase (AST), - Blood urea nitrogen, - Gamma glutamyltransferase (GGT).

[0019] According to another object, a computer program product is disclosed comprising code instructions for implementing the method according to the above description when the instructions are executed by a processor.

[0020] In accordance with another object, a computing system is disclosed, which comprises: - processor, - a non-transitory computer-readable medium storing program code executable by a processor; The processor is configured to execute program code for performing operations including applying a trained model on a set of values ​​of biological variables for a subject, the trained model being configured to predict the chronological age of the subject based on these sets of values ​​to obtain a predicted age of the subject, the predicted age of the subject corresponding to a physiological age.

[0021] In an embodiment, the processor is further configured to calculate a difference between the physiological age of the subject and a reference age corresponding to an average age predicted by the trained model on a population, the population including a plurality of individuals of the same chronological age as the subject, or a plurality of individuals of different chronological ages within a chronological age range of at least 50 years.

[0022] In an embodiment, the processor is communicatively coupled to the client system via a data network and configured to receive a set of values ​​of the biological variables relating to the subject from the client system and return to the client system the physiological age of the subject or the difference between the physiological age of the subject and a reference age corresponding to the average age predicted by the trained model on a population.

[0023] In an embodiment, if the calculated difference is different from zero, the processor is further configured to calculate a SHAP value associated with each of the biological variables and identify the SHAP value having the highest absolute value, said SHAP value corresponding to the biological variable that contributes most to the calculated difference.

[0024] In an embodiment, the processor is further configured to generate graphical data representing the SHAP values ​​having the highest absolute values, where the SHAP values ​​that contribute to increasing the age predicted by the model relative to the reference age are represented in a first color and the SHAP values ​​that contribute to decreasing the age predicted by the model relative to the reference age are represented in a second color.

[0025] In an embodiment, the computing system further comprises a memory, and the processor comprises: - calculating, for each individual in a population of individuals for a plurality of chronological ages, a predicted age for the individual based on a set of values ​​of biological variables for the individual; - calculating, for a plurality of chronological ages of individuals of the population, a mean value of each biological variable for individuals for said chronological ages whose predicted age is equal to the chronological age, and storing said mean values ​​in a memory; - o at least one value of a biological variable that contributes most to the difference between the subject's predicted age and the reference age; ○ The reference values ​​of the biological variables for the same chronological age To calculate the difference between Further configured.

[0026] Description of the drawings Other characteristics and advantages of the invention will become apparent on reading the following detailed description, given by way of non-limiting example, and on referring to the accompanying drawings, in which: [Brief description of the drawings]

[0027] [Figure 1] FIG. 2 is a schematic diagram illustrating the main steps of a method according to one embodiment. [Diagram 2] 1 is a schematic diagram illustrating a computing device according to one embodiment. [Diagram 3] 3a and 3b are plots of the performance of an XGBoost model for predicting chronological age on the training and validation datasets, respectively, and FIG. 3c and 3d are plots of the performance of an XGBoost model using a custom loss gradient function as a function of chronological age on the training and validation datasets, respectively. [Figure 4] 1 is a chart showing the relative importance of the 20 most important variables in physiological age without contextualization. [Diagram 5] FIG. 1 shows a series of contextualized partial dependence plots for multiple biological variables. Each point represents an individual and the shade represents chronological age. On the x-axis are the real values ​​of the variables and on the y-axis are the SHAP values ​​given to this individual for this variable. [Figure 6] 6a and 6b show the clustering of the contextualized SHAP values ​​calculated in the NHANES study and the mean SHAP value of the most important biological variables for each cluster. [Figure 7] FIG. 13 depicts an exemplary display of personal results including the most important contextualized SHAP values ​​contributing to the difference between physiological age and predicted age. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0028] Detailed Description of the Preferred Embodiments Now, with reference to the drawings, a computer-implemented method for determining the physiological age of a subject will be described.

[0029] Computing Systems This method for determining the physiological age of a subject may be implemented by a computing system 1 as shown generally in FIG. 2, comprising at least one processor 10, which may include one or more computer processing units (CPUs) and / or graphic processing units (GPUs), and a non-transitory computer-readable medium 11 storing program code executable by the processor to implement the method described below.

[0030] The computing system 1 may also include at least one memory 12 that stores a trained model configured to predict the chronological age of a subject based on a plurality of biological variables. The memory 12 may be the same as or different from the non-transitory computer-readable medium 11 that stores the program code. The memory may be, for example, a random access memory (RAM), a magnetic hard disk, a solid-state disk, an optical disk, an electronic memory, or any type of computer-readable storage medium.

[0031] The memory 12 may also store other reference data obtained by application of the trained model on a reference population and used as a reference in the below detailed steps of the method. For example, the memory 12 may store the mean age predicted by the model on a reference population including a plurality of individuals for a plurality of chronological ages.

[0032] The memory also stores, for each of the plurality of chronological ages: - a mean age predicted by the model on a reference population including a plurality of individuals of said chronological age; - Reference values ​​of a number of biological variables relative to said chronological age. In an embodiment, the computing system 1 may be communicatively coupled to a client system via a data network 3, e.g., a wireless network. The client system may be a computing system located in a medical facility such as a hospital, laboratory, clinic, etc. One of the computing system 1 and the client system 2 may include a screen 4 for displaying relevant data obtained through implementation of the method.

[0033] The same or a distinctly different computing system, also comprising at least one processor and a non-transitory computer-readable medium 11, may also be used to train a model on the normative database 13 to predict the chronological age of the subject.

[0034] Method for determining physiological age of a subject and detecting premature aging With reference to FIG. 1, a method for determining the physiological age of a subject may include, for the subject under consideration, a preliminary step 90 of receiving a set of values ​​of biological variables relating to the subject.

[0035] In embodiments, the biological variables are laboratory available variables, i.e., variables that can be obtained in a laboratory, for example by blood analysis, urine analysis, saliva analysis, other biological fluid analysis, or direct measurement on the subject during a clinical test. In embodiments, the method may further include receiving socioeconomic or sociodemographic variables for the subject in addition to the biological variables.

[0036] The biological variables may include at least one variable, the at least one variable being glycated hemoglobin.

[0037] The biological variables may include at least one variable, and preferably several, for example at least five, or all, of the following groups: - glycated hemoglobin, - Creatinine in urine, - Cholesterol, - Alanine transaminase (ALT), - mean corpuscular volume, - Aspartate transferase (AST), - Blood urea nitrogen, - Gamma glutamyltransferase (GGT).

[0038] In an embodiment, the biological variables may further include at least one additional variable, preferably a plurality, for example at least five, or all, of the following groups: -Lynn, - Triglycerides, - Albuminemia, - serum glucose, - Red blood cell distribution width, - serum folate, - creatinine, - alkaline phosphatase (ALP), - hematocrit, - Albuminuria, - Osmolality, - C-reactive protein, - Lymphocyte count.

[0039] In an embodiment, the biological variables may further include at least one additional variable, preferably a plurality, for example at least five, ten, or all of the variables listed below. - Mean corpuscular hemoglobin concentration (MCHC), - mean corpuscular hemoglobin, - monocyte count, - monocyte percent, - red blood cell count, - Folic acid, RBC - ferritin dose, - Direct HDL-cholesterol, - Basophil count, - Basophil percent, - bicarbonate, - hematocrit, - Total bilirubin, - Potassium, - Segmented neutrophils num, - percent segmented neutrophils, - sodium, - Total calcium, - total protein, - uric acid, - White blood cell count.

[0040] A person skilled in the art may refer to NHANES laboratory methods, such as NHANES 2017-2020 laboratory methods, for methods for assessing each of the above biological variables.

[0041] The values ​​of the biological or physiological variables may have been obtained from the subject and stored in a memory. Receiving the set of values ​​may then include receiving the data over a data network or accessing the memory in which they are stored for further processing. Receiving the set of values ​​may also include the computing system 1 receiving said set of values ​​from the client system 2 over a data network. The set of values ​​may be transferred in an encrypted manner or over a secure channel.

[0042] The method then includes applying the trained model configured to predict the chronological age of the subject from said set of values ​​on the set of values ​​of the biological variables to obtain a predicted age for the subject 100. The predicted age, determined based on the set of values ​​of the biological variables, corresponds to the physiological age of the subject, which may be equal to the chronological age of the subject, or may be lower or higher than the chronological age of the subject. In the latter case, it corresponds to premature aging of the subject, since the physiological state of the subject implies that it is older than its chronological age.

[0043] The method may therefore include comparing the subject's predicted age with its chronological age 110 and inferring that if the difference between the physiological age and the chronological age is positive, the subject is at increased risk of premature aging and developing chronic diseases, such as diabetes, coronary heart disease, or kidney disease.

[0044] In other embodiments, the method may include comparing 120 the subject's predicted age with a reference age (which may be stored in memory 12) corresponding to the average age predicted by the trained model on a population of the same chronological age as the subject, and inferring premature aging of the subject and an increased risk of developing chronic diseases, such as diabetes, coronary heart disease, and kidney disease, if the difference between the subject's predicted age and the reference age is positive.

[0045] The computing system may return to the client system 2 in sub-step 130 the physiological age obtained for the subject and / or the difference between the physiological age and the reference age.

[0046] In an embodiment, the physiological age obtained for a subject can be monitored at different time points to track the progression of the physiological age of said subject. Age progression is natural and is monitored periodically to track changes in the subject's health (e.g. to detect harmful abnormalities and conduct research into their causes in order to prevent or treat them) or to study the influence of parameters on aging (e.g. treatment, anti-aging treatment, infectious diseases, chronic diseases, chronic disease treatment, diet, physical or mental stress). In particular, harmful abnormalities are detected when the difference between the predicted age and the reference age of the subject is positive. Thus, in an embodiment, the physiological age of a subject is calculated at least twice to track the progression of the physiological age of said subject.

[0047] Training a predictive model The model is pre-trained by supervised learning on a training dataset that includes, for a plurality of individuals in the population, the values ​​of each individual's chronological age and an initial set of biological variables, which may include, for example, some or all of the biological variables listed above.

[0048] In an embodiment, the initial set of biological variables includes at least 10 variables, and preferably at least 20 variables.

[0049] The population preferably includes individuals of chronological age within a broad age range, preferably at least 50 years, without a large imbalance in the number of men and women across the age groups. The population may include more than 1,000 individuals, preferably more than 10,000 individuals.

[0050] The training dataset is split into a training subset (approximately 80%) and a validation subset (approximately 20%). Model training is performed to minimize the mean absolute error (MAE) between the age predicted by the model and the subject's chronological age.

[0051] The trained model preferably includes:

number

[0052] Here, grad i is the gradient to be calculated for the i-th individual,

number

[0053] Such a custom loss function as a function of chronological age allows to moderate the bias of the model predicting younger and older, respectively. Furthermore, the choice of XGboost model allows the management of missing data and enables the interpretability of the model.

[0054] Training the model may also include using a feature selection algorithm, for example the GrootCV algorithm, to eliminate an initial set of variables whose contribution is not statistically greater than chance.

[0055] In addition, recursive feature elimination (RFE) may be implemented to remove variables with minimal contributions such that the removal does not impair the quality of the model, in which case the set of values ​​of the biological variables used to determine the physiological age of the subject comprises one value for each biological variable retained at the end of the feature selection.

[0056] 1, once the physiological age of the subject has been determined, or when a positive difference between the physiological age of the subject and the chronological age, or between the physiological age of the subject and the average predicted age for a population of individuals with the same chronological age, has been calculated, the method may further include determining 200 the contribution of each variable to the age predicted by the model and identifying the biological variables that contribute most to the difference. This step may include identifying a predetermined number, for example 10 or less, for example 5, of the variables that contribute most to the difference.

[0057] This can be done by calculating SHAP values ​​for all biological variables used to determine physiological age. SHAP stands for Shapley Additive exPlanations, and SHAP values ​​were originally proposed by Lundberg, Scott et al. in "Consistent individualized feature attribution for tree ensembles," 2019.

[0058] The sum of the SHAP values ​​for all biological variables in the model represents the individual deviation from the norm.

[0059] In the first embodiment, the standard is the average age predicted by the model over the entire data set.In the experiment detailed below, where the data set is the NHANES data set, the average predicted age over the entire population is 39.9 years old.Therefore, for a given subject, physiological age is the average age predicted by the model over the entire data set plus the sum of all the SHAP values ​​of all biological variables, respectively.

[0060] In a second embodiment, the reference is the average age predicted by the model over a sub-portion of the data set that includes only individuals of the same chronological age as the individual. For a given subject, the physiological age is then the average age predicted by the model for a population that includes only individuals of the same chronological age plus the sum of all SHAP values ​​of all biological variables, respectively. In this case, the SHAP value is denoted as contextualized. The sum of the contextualized SHAP values, hereafter denoted as "iCAD", thus represents the difference between the physiological age of the subject and the average physiological age of the population of the same chronological age. A positive sum corresponds to an increased risk of premature aging and death for the subject.

[0061] In both embodiments, the method may include determining 200 the SHAP values ​​associated with each of the biological variables and identifying those with the highest absolute values, in particular the positive SHAP values ​​with the highest values, since these correspond to the biological variables that contribute most to the increase in physiological age with respect to the baseline.

[0062] Referring to FIG. 7, the computing system 1 may generate, in step 210, graphical data to be transmitted to the client device 2 and displayed on the screen 4, the graphical data representing SHAP values ​​explaining the difference between the predicted physiological age for the subject (in the figure, f(x)=58.884 years) and the reference age (E[f(X)]=45.29 years), which is the average predicted age of a population of individuals of the same chronological age as the subject in FIG. 7. The display of the SHAP values ​​may include a chart, preferably in increasing order of importance from bottom to top, in which the horizontal axis represents age and the vertical axis represents the biological variables that contribute most to the difference between the physiological age and the reference value and their corresponding SHAP values. Each SHAP value may be represented by an arrow whose length is in accordance with the scale of the horizontal axis, with positive SHAP values ​​being shown in a first color and negative SHAP values ​​being shown in a second color. Also, the direction of the arrow is determined according to the sign of the SHAP value, since negative SHAP values ​​tend to lower the predicted age and positive SHAP values ​​tend to raise the predicted age.

[0063] Additionally, after the biological variables that contribute most to the difference between physiological age and the reference value have been identified, the subject may undergo regular surveillance of at least one of the variables that contribute most to the difference.

[0064] In an embodiment, once the biological variable that contributes most to the difference between physiological age and reference age has been identified, in step 300, the corresponding value for the subject may be compared to the reference value of said biological variable for the same chronological age as the individual.

[0065] Reference values ​​for each biological variable and for each physiological age may also be established prior to implementing the method for determining the physiological age of a subject using the predictive model and its training data set and may be stored in memory 12.

[0066] Referring to FIG. 5, a contextualized partial dependence plot is shown for multiple biological variables including glycated hemoglobin, urinary creatinine, blood urea nitrogen, mean corpuscular volume, cholesterol, triglycerides, red blood cell distribution width, and phosphorus. Each plot displays multiple dots, each dot representing one individual, the horizontal axis represents the value of the corresponding biological variable, and the vertical axis represents the contextualized SHAP value of said biological variable. The gradation of the dots represents the chronological age of the individual. Thus, it can be noticed that according to the gradation, the value of the biological variable with a SHAP value equal to 0 changes according to age.

[0067] Therefore, the chronological age reference value for each biological variable can be determined as the average value of the biological variable for which the corresponding SHAP value of the biological variable is zero, i.e., the biological variable does not contribute to the difference between predicted age and chronological age.

[0068] In another embodiment, a chronological age reference value for each biological variable may be determined as the average value of the biological variable among a plurality of individuals of said chronological age whose predicted age is less than or equal to said chronological age.

[0069] In an embodiment, when the biological variables that contribute most to the difference between the physiological age and the reference age have been identified, the method may also include determining 400 an aging profile of the subject from said biological variables from a plurality of pre-defined clusters, each cluster corresponding to an aging profile, and the clusters are established based on the biological variables that contribute most to the difference between the physiological age predicted by the model and a common reference value for a population including a plurality of individuals within a plurality of chronological age ranges. Preferably, the population includes a plurality of individuals each of a plurality of chronological ages spanning an age span of at least 50 years.

[0070] More specifically, the clusters are: - implementing a trained model for predicting the chronological age of a plurality of individuals of a population, thereby obtaining a physiological age for each individual; - determining at least an average predicted age across the population, which may be a single average predicted age across the population or which may include, for each of a plurality of chronological ages, a predicted age of a subset of individuals in the population at said chronological age; - determining, for a plurality or all of the individuals of the population, the SHAP values ​​of the biological variables that contribute most to the difference between the predicted age for the individual and the mean predicted age, this step may comprise, for example, determining the 10 or 20 highest SHAP values ​​in absolute value; - performing clustering on the SHAP values ​​obtained for multiple individuals.

[0071] Clustering may be performed by applying a clustering algorithm to the SHAP values, for example an agglomerative clustering algorithm, such as Ward's algorithm and Euclidean distance for linkage.

[0072] The method may further include generating a graphical representation of the obtained clusters, which may include applying an algorithm for reducing the dimension of the SHAP values ​​and displaying a 2D representation of the clusters by associating each dot corresponding to a cluster with a respective color. The dimension reduction may be performed, for example, by Uniform Manifold Approximation and Projection (UPAM) or Principal Component Analysis.

[0073] With reference to Figure 6a, a graphical representation of the clustering of SHAP values ​​obtained in the NHANES dataset (see below) is illustrated, which allows the identification of 10 clusters. In Figure 6b, an average individual representative of each cluster is illustrated and, starting from the bottom, the cumulative contribution of each contextualized SHAP value is presented (in positive and negative values) relative to the predicted final value at the top of the figure.

[0074] It can therefore be seen that the population can be clustered into different groups depending on the respective contribution of different biological variables to the difference between predicted age and reference value. EXAMPLES

[0075] Working Example NHANES Dataset A consistent and comprehensive dataset was constructed in three steps: (i) All NHANES data from 1999 to 2018 were merged, resulting in 36,945 variables. (ii) Laboratory variables were selected and aggregated using a dedicated web interface. (iii) The largest data set corresponding to the inclusion criteria, including the minimum amount of missing data, was defined.

[0076] The final dataset included 48 laboratory variables (Table S1) for 60,322 individuals, including 30,747 females and 29,575 males, with mean ages of 39.3 ± 19.7 and 39.5 ± 20.2 years, respectively. The amount of data for 12-20 years was twice that of other ages, with a 25% decrease for available subjects aged 70-79 years. No major imbalance in the number of men and women was noted across age groups. The amount of missing data was low (25% of individuals with one missing value accounted for 0.06% of the total values) and evenly distributed in age and gender. These were mainly related to missing data for C-reactive protein, folate, albumin, and creatinine. In a next step, imputation methods for missing data were implemented, except for XGBoost, which can natively manage missing data.

[0077] [Table 1A] [Table 1B]

[0078] Selecting the best explainable algorithm for defining personalized physiological age (PPA) To define the best explainable predictive algorithm to define PPA, different machine learning algorithms were evaluated using training and test datasets corresponding to 80% and 20% of the original dataset. To reduce the number of variables and the expected overfitting of the model, GrootCV feature selection was used to eliminate variables whose contribution was not statistically greater than chance. Four variables were eliminated: basophil count, mean corpuscular hemoglobin, monocyte count, and percent segmented neutrophils, which reduced the number of variables to 44. The choice of whether to keep a variable was based in part on redundancy. A variable was retained alone to contribute to PPA when the performance of another biologically linked parameter was comparable or better. Five classes of machine learning algorithms were then compared to predict chronological age: tree-based models (decision trees, random forests, and XGBoost), regularized regression methods (ElasticNet, a method that includes both L1 and L2 norm regularization of the coefficients), and neural networks (multilayer perceptron, MLP).

[0079] For each model (Table S2), a grid search exploration of hyperparameters with 5-fold cross-validation was performed using the training dataset. Models were run in R 2 The performance of the algorithms was evaluated based on the results on the test dataset using the coefficient of determination (Coefficient of Determination) and the mean absolute error (MAE). 2Similar performance was observed on the training and testing datasets for both sigma-beta and MAE. XGBoost and MLP (Multilayer Perceptron) achieved the best similar performance, with XGBoost also showing the lowest standard deviation in cross-validation. Due to the high dimensionality (large number of variables) and number of subjects in the database, XGBoost was selected as the model for its fastest explainability calculation. Error analysis revealed a differential bias in the model predicting age, which tended to predict younger individuals as older and older individuals as younger (Fig. 3a, b). To correct for the bias, a custom objective function was introduced during training of XGBoost, which significantly minimized the bias while maintaining performance (Fig. 3c, d) (R 2 and MAE on the test dataset (0.72 and 8.1, respectively).

[0080] · Explainability of physiological age To define the contribution of each variable to the individual PPA prediction, the Shapley Additive exPlanations (SHAP) TreeSHAP framework was used in an XGBoost model with a custom loss model. The sum of the SHAP values ​​of all variables in the model represents the individual deviation from the average predicted chronological age across the entire dataset (39.9 years in our model, i.e., the base value). For a given individual, the predicted age was 39.9 years plus the sum of all SHAP values. For a set of variables, the higher the overall SHAP value, the more the variable contributes positively to PPA.

[0081] Ranking was performed by the average absolute value of the global SHAP contribution for each variable. From the top 20 variables, many variables were related to metabolism, whether it was nitrogen-related (e.g., urinary metabolites, creatinine), carbon-related (e.g., glycohemoglobin, triglycerides, glucose), or liver function-related (e.g., albumin, ALT, GGT). Glycohemoglobin emerged as the parameter with the highest contribution (10.7% of the mean total SHAP contribution), while serum glucose was ranked 9th. Urine and blood creatinine, reflecting renal function, were also shown to contribute to PPA prediction. Several parameters directly or indirectly related to red blood cells, mean corpuscular volume, red blood cell distribution width, hematocrit, and serum folate were also distributed among the top 20 variables. Immunity / inflammation related traits (C-reactive protein and lymphocyte count) ranked 19th and 20th respectively, while other parameters related to the immune system (e.g. monocyte or lymphocyte percentage, white blood cell count) had a lower impact on the SHAP value. The age trends of their mean values ​​usually follow the SHAP value (positive or negative). For example, the mean raw value of glycated hemoglobin increases with age, as do higher raw values ​​with higher SHAP values. For the majority of variables (11 out of 20 variables), the higher the variable value, the greater the deviation from chronological age. No obvious changes in the explanatory profile were found between men and women with similar rankings of variables.

[0082] · PPA contextualization explainability by age group. The principle of contextualization is to provide a better explanatory model by taking the average predicted value of individuals sharing the same chronological age as the base value (instead of the average predicted value for the whole population). The SHAP contribution of each variable is then, therefore, called the "contextualized SHAP". Glycated hemoglobin, blood urea nitrogen, mean corpuscular volume, and urinary creatinine were found to have consistent contributions throughout the lifespan, although there is a stronger contribution between 40 and 70 years of age. Other variables had more age-specific contributions, e.g. alkaline phosphatase (12-18 years), ALT and cholesterol (20-40 years), or lymphocyte count and folate (60 years and older). We derived the "iCAD" metric, which was defined as the sum of the contextualized SHAP values ​​for a given individual.

[0083] · iCAD validation and robustness Using multivariate Cox survival models, iCAD was found to be a relevant predictor of mortality (Table 1). Adjusted hazard ratios for sex, chronological age, and year of participation showed that negative iCAD values ​​were associated with a reduced risk of mortality, but non-significantly (aHR with 95% Cl of 0.88 [0.76;1.03] for the 1st decile compared with the 5th decile taken as reference). Positive iCAD values ​​were significantly associated with a progressively increased risk of mortality (aHR 95% CI 1.18 [1.01;1.38], 1.37 [1.17;1.59], 1.38 [1.18;1.60], and 1.69 [1.45;1.97] for the 7th to 10th deciles, respectively).

[0084] [Table 2]

[0085] Partial dependence of contextualized SHAP values ​​as a new PPA metric Partial dependence shows that the contextualized SHAP contribution to the PPA prediction changes according to the variation of the raw variable values ​​(Figure 5). This relationship for a given variable (shape of the curve) seemed to be quite similar among the ages, although the amplitudes were different. It is also possible to notice different types of relationships, for example a sigmoidal rise (e.g., glycosylated hemoglobin, blood urea nitrogen), a sigmoidal fall (e.g., phosphorus), or a linear trend (e.g., folate, urinary creatinine). These profiles clearly showed different ranges of variable values ​​for which the corresponding contextualized SHAP value was positive, neutral, or negative. For example, the contextualized SHAP value was negative at low values ​​for glycosylated hemoglobin, but a sharp increase appeared in the range of values ​​between 5 and 6%. This transition zone was characterized by a passage starting from zero and differed according to age. Thus, a threshold of 5.4% seemed to characterize the “normal” range for younger subjects, but changed with age, increasing to 5.8% for subjects over 50 years old. For urinary creatinine, the SHAP contribution decreases as a result of increasing values, with the null SHAP value being around 10,000 μmol / L. Figure 5 clearly shows the decrease in the normal range of values ​​with age.

[0086] Biological parameters and aging To identify the supposed specific features that are the origin of the profiles for an individual, all contextualized SHAP values ​​were clustered, regardless of chronological age (Fig. 6a,b). The clustering highlights 10 SHAP clusters grouped into two classes according to glycated hemoglobin SHAP values. The contribution of low glycated hemoglobin (below the clinical threshold at 6%) appeared to correlate with a "low" physiological age in older individuals, as in cluster 2. Changes in a reduced set of variables, including creatinineuria, cholesterol, ALT, mean corpuscular volume (MCV), AST, blood urea nitrogen (BUN), and GGT, differentiated the clusters within each class. All other variables contributed slightly to the differences between the clusters. Fig. 6b shows the profiles of the variable SHAP values ​​for each cluster. This suggests that different profiles corresponding to the same iCAD may reflect different physiological aging modalities. Clusters 2 and 4 were characterized by systematic negative and positive deviations of key biological variables according to negative and positive deviations from chronological age. All other profiles were characterized by a mixture of positive and negative SHAP values ​​for the significant variables.

[0087] · Generate minimal models by recursive feature elimination In terms of therapeutic use, the best compromise had to be found between the accuracy of the PPA estimation and the minimum number of relevant features. The results of the exhausted RFE algorithm show that 26 variables are required for the R 2 It was shown that the proposed method is sufficient to predict PPA without significantly degrading the performance of the model estimated by

[0088] [Table 3] [Explanation of symbols]

[0089] 1. Computing Systems 2 Client System 3. Data Network 4 Screens 10 Processors 11 Non-Transitory Computer-Readable Media 12. Memory 13 Reference Database

Claims

1. 1. A computer-implemented method for determining a physiological age of a subject, comprising: applying (100) a trained model on a set of values ​​comprising at least values ​​of a biological variable for the subject to obtain a predicted age of the subject, the trained model being configured to predict a chronological age of the subject based on the set of values; The computer-implemented method, wherein the physiological age corresponds to the predicted age.

2. 2. The computer-implemented method of claim 1, further comprising comparing the physiological age of the subject to the chronological age of the subject, wherein a positive difference between the physiological age and the chronological age of the subject indicates premature aging of the subject.

3. 2. The method of claim 1, further comprising: comparing the physiological age of the subject to a reference age corresponding to an average age predicted by the trained model on a reference population; and identifying (200) the biological variables that contribute most to the difference when the physiological age of the subject differs from the reference age.

4. 4. The computer-implemented method of claim 3, wherein the reference population is a population of individuals having the same chronological age as the individual.

5. 5. The computer-implemented method of claim 3, wherein identifying the biological variables that contribute most to the difference comprises determining a SHAP value associated with each of the biological variables and identifying the SHAP value having the highest absolute value.

6. 6. The computer-implemented method of claim 3, further comprising a step of comparing (300) at least one value of the biological variable that contributes most to the difference with a reference value of said biological variable for the same chronological age.

7. 7. The computer-implemented method of claim 6, wherein the reference value of a biological variable for a given chronological age is determined as an average value of the biological variable among a plurality of individuals of the given chronological age in which the biological variable does not contribute to a difference between the predicted age and the chronological age.

8. 8. The computer-implemented method of claim 3, further comprising: determining (400) an aging profile for the subject from among a plurality of pre-established aging profiles based on the identified biological or physiological values ​​that contribute most to the difference.

9. The plurality of pre-established aging profiles comprises: - predicting the chronological ages of a plurality of individuals of a population by implementing the trained model, the population including a plurality of individuals for each of a plurality of chronological ages; and - determining at least a mean predicted age of said population; - determining, for a plurality of individuals of the population, the SHAP values ​​of the biological variables that contribute most to the difference between the predicted age of the individual and the mean predicted age; - performing clustering on the SHAP values ​​to obtain a finite number of clusters, each cluster corresponding to an aging profile.

10. 10. The computer-implemented method of claim 1, wherein the trained model is an XGboost model that uses a custom loss function that is a function of chronological age.

11. A computer program product comprising code instructions for implementing a method according to any one of claims 1 to 10, when said instructions are executed by a processor (10).

12. a processor (10); - a non-transitory computer-readable medium (11) storing program code executable by said processor, 1. A computing system (1), wherein the processor is configured to execute the program code for performing operations including applying a trained model on a set of values ​​of biological variables for the subject, the trained model being configured to predict the subject's chronological age based on the set of values ​​to obtain a predicted age for the subject, the predicted age for the subject corresponding to a physiological age.

13. 13. The computing system (1) of claim 12, wherein the processor (10) is communicatively coupled to a client system (2) via a data network (3) and configured to receive from the client system the set of values ​​of biological variables relating to the subject and to return to the client system the physiological age of the subject or a difference between the physiological age of the subject and a reference age corresponding to an average age predicted by the trained model on a population.

14. 14. The computing system of claim 13, wherein if the calculated difference is different from zero, the processor is further configured to calculate a SHAP value associated with each of the biological variables and to identify the SHAP value having the highest absolute value, the SHAP value corresponding to the biological variable that contributes most to the calculated difference.

15. 15. The computing system of claim 14, wherein the processor (10) is further configured to generate graphical data representing the SHAP values ​​having the highest absolute values, the SHAP values ​​contributing to increasing the age predicted by the model with respect to the base age being represented in a first color and the SHAP values ​​contributing to decreasing the age predicted by the model with respect to the base age being represented in a second color.

Images