Method for calculating biological age

Abstract

The invention relates to a method for calculating biologica! age of a person based on artificial intelligence, which involves the analysis of predefined cells from the human body.

Description

[0001] " METHOD FOR CALCULATING BIOLOGICAL AGE"

[0002]

[0003] FIELD OF THE INVENTION

[0004] The present invention relates to a method for calculating a person's biological age. In particular, the invention relates to a method for calculating the biological age based on the fatty acid quantitative values of the erythrocyte membrane phospholipids of people and with the aid of artificial intelligence.

[0005] BACKGROUND OF THE INVENTION

[0006] The calculation of biological age has an important place in ageing research, as well as its potential use to highlight the connections with risk factors in subjects and apply preventive measures to slow down an accelerated ageing process.

[0007] There are several known methods for estimating biological age, including multiple linear regression (MLR), principal component analysis (PCA), Hochschild method, and Klemera and Doubal method (KDM) (Jia et al, Clin. Interv. in Aging 2017, 12, 759-772). These methods are based on calculations that take into account chronological age and specific biomarkers of aging.

[0008] The fundamental differences between established aging clocks lie in the role of chronological age and the specific selection criteria for biomarkers. Chronological age, the literal number of years since birth, is influenced by genomic and epigenomic factors, including environment, nutrition, and lifestyle, which modulate the effects of time on a subject to determine his / her biological age. Biological age uses the parameter of "years" to express an individual’s cumulative damage (or transformation) and its trajectory over time.

[0009] To establish the "bias" between chronological years and real-world physiological effects, the choice of biological site and molecular parameters is crucial. This allows for an age assessment expressed as the sum of YEARS + METABOLISM + EXPOSURE + RESPONSE. Current methods for calculating biological age rely on population-scale parameters. Common approaches include analyzing the state of genetic material, specifically epigenetic modifications like cytosine methylation (Zhao W, Ammous F, Ratliff S, Liu J, Yu M, Mosley TH, Kardia SLR, Smith JA. Education and lifestyle factors are associated with DNA methylation clocks in older African Americans, Int. J. Environ. Res. Public Health 2019, 16, 3141. 10.3390 / ijerph16173141) or measuring telomere shortening(Marioni, RE et al. Int. J. Epidemiol. 2018, 47, p. 356, 10.1093 / ije / dyx233;447(7147):924- 931; Sanders JL, Newman AB. Telomere length in epidemiology: A biomarker of aging, age-related disease, both, or neither? Epidem. Rev. 2013, 35, 112-131).

[0010] At the metabolic level, "omics" technologies provide high-resolution data from various sources: transcriptomics provides accuracy near the theoretical limit (Meyer, DH; Schumacher, B. BiT age: A transcriptome based aging clock near the theoretical limit of accuracy. Aging Cell. 2021, 20, e13320), while the taxonomic status of the microbiome (gut, skin, or oral, Galkin, F.; Mamoshina, P.; Aliper, A.; Putin, E.; Moskalev, V.; Gladyshev, VN; Zhavoronkov, A. Human Gut Microbiome Aging Based on Taxonomic Profiling and Deep Learning iScience 2020, 23, 101199; Human Skin, Oral, and Gut Microbiomes Predict Chronological Age. mSystems 2020, 5, e00630-19) can predict age with high specificity. Other "clocks" utilize multimodal data, ranging from brain MRI and retinal imaging images (Wang et al. Accurate estimation of biological age and its application in disease prediction using a multimodal image Transformer system. PNAS 2024, 121, e2308812120) to standard clinical laboratory markers (Erema, VV, et al. Int. J. Mol. Sci. 2022, 23, 15103; see figure). Recently, twelve "hallmarks of aging" — including genomic instability, mitochondrial dysfunction, and dysbiosis — were proposed to provide a more comprehensive diagnostic framework (Lopez-Otin, C. et al. Hallmarks of aging: an expanding universe. Cell 2023, 186, 243-278). However, a primary requirement for any biological clock remains its viability for "real-world" clinical application without loss of diagnostic value.

[0011] While blood is a common sample for metabolomics, it is essential to distinguish between metabolites dissolved in plasma (e.g., sugars, enzymes, circulating lipids) and those which are building blocks of cellular compartments. This differentiation is particularly important for the lipid molecules, and in particular for the fatty acidbuilding blocks that are found in many lipid molecules with different functions. For example, palmitic acid in plasma can be found in triglycerides, endocannabinoids, and other circulating lipids. On the other hand, it can also be found as one of the saturated fatty acid moieties of phospholipid molecules that assemble with other types of phospholipids, containing a variety of saturated, monounsaturated and polyunsaturated fatty acids, to form the cell membrane. Inthis latter case, palmitic acid - as well as the other fatty acid components of phospholipids - provides completely different information which has, at the same time, metabolic and structural significance. A meaningful example of this significance is the role of palmitic acid-based phospholipids, which affect membrane rigidity and function, and become a specific indicator of changed biophysical characteristics in patients having a precise health status like obesity (Sot, J.; et al. Erythrocyte Membrane Nanomechanical Rigidity Is Decreased in Obese Patients. Int. J. Mol. Sci. 2022, 23, 1920). Therefore, the use of lipids as biomarkers in whole blood is not equivalent to the membrane lipidome information, despite the identical nomenclature of the molecules. This distinction between blood and cellular samples (such as erythrocytes, the most abundant cells in blood) is critical for accurate diagnostic reporting and for the use of the information described in this patent.

[0012] In patent application WO2012164525, human plasma is used to analyse molecular levels of circulating lipids and fatty acids, acylcamitines, sugars, etc., and to calculate the biological age of individuals. It is clear that this information is unrelated to the specific process or to membrane compartment formation and functions. Moreover, it is well known that plasma / serum lipidome is a short-term information. The most stable information is embedded in the membrane lipidome, specifically in red blood cells (RBCs, erythrocytes), given their 4-month half-life, compared with short-lived blood cells, such as lymphocytes and platelets. RBCs are well-known as good representative cells for long-term evaluation of health status and cellular functions. Such differences, between the biological significance of samples like blood and erythrocytes, are clearly indicated for diagnostic purposes (Brenna, J. T., Plourde, M., Stark, K. D., et al. Best practices for the design, laboratory analysis, and reporting of trials involving fatty acids. Am. J. Clin. Nutr. 2018, 108, 211). The importance of the RBC membranes has been largely studied in membrane lipidomics for human health (C. Ferreri, C. Chatgilialoglu, Membrane Lipidomics for Personalized Health, Wiley 2015), under different physiological and pathological conditions, evidencing that some fatty acids in the membrane phospholipids can be used as biomarkers of the onset or progress of pathologies. For example, the “omega-3 index”, calculated from the sum [%EPA (omega-3) + % DHA (omega-3)] in RBC membranes, has beendeveloped as a cardiovascular risk index (Harris, William S. Recent studies confirm the utility of the omega-3 index. Curr. Opin. Clin. Nutr. Metab. Care 2025, 28, 91-95). Actually, it is becoming increasingly evident that the lipidome is a more sensitive indicator of "real-time" health than the genome, as it reflects the combined impact of diet, environment, and metabolism on cellular integrity. On this basis, a new concept of Biological Age can be envisioned using the membrane lipidome of RBCs and the fatty acid moieties of their phospholipids. RBCs are exposed throughout their 4-month lifetime to endogenous and exogenous “stressors” and are representative of what happens to the cell membrane structures of all other cells in the human body. Moreover, cell membranes have a constant composition for each tissue, which is called the “lipid code”, to be maintained through an efficient turnover and support the appropriate organs’ functions (Harayama, T., Riezman, H. Understanding the diversity of membrane lipid composition. Nat Rev Mol Cell Biol 2018, 19, 281-296). The turnover of membrane phospholipids is driven by the recycling of modified / oxidised phospholipids, a fundamental, enzymatically driven, and signalling-controlled process (called Lands’ cycle). On these premises, it is possible to envisage, using the membrane lipidome composition, the measure of the body’s exposure to insults and to identify biological checkpoints to maintain favourable homeostatic conditions. This homeostasis occurs in different ways for each individual; in the most successful case, it can slow all ageing processes, reduce inflammatory processes (which are responsible for ageing: inflammaging), and enhance protective signalling cascades, which together preserve the body from chronological ageing. It is also important to consider that, nowadays, the RBC membrane lipidome check-up at its mature age (4 months) is available through an accredited analytical protocol in accordance with ISO 15189, thereby providing high-precision, diagnostic value of these data.

[0013] Methods for establishing biological age must describe not only the biological sample (blood) but also the cell of interest (red blood cell) and the specific molecular components (membrane phospholipids) from which parameters for calculating biological age are obtained. At the moment, the DNA target appears to be the most attractive, also because without DNA, there is no form of life. On the other hand, cells cannot exist without the cell membrane, and a method to evaluatethe age using information from this compartment has its own biological value. A methodological problem in the biological age assessment procedure remains to be resolved: exposure to DNA epigenetic factors cannot be considered exhaustive in assessing all the phenomena that influence chronological age, also due to the location of DNA in the nucleus, or in a "separate" area of the cellular organization.

[0014] To develop a biological age calculation procedure that is more effective at expressing exposure over the life span, to overcome the drawbacks of the known art, and to achieve these and further purposes and advantages, the Applicant has studied, tested, and created this invention.

[0015] SUMMARY OF THE INVENTION

[0016] The present invention is set forth and characterized in the independent claim. The dependent claims describe other characteristics of the present invention or variants of the main solution idea.

[0017] In accordance with the aforesaid objectives and in order to solve the above mentioned technical problem in a novel and original manner, obtaining also relevant advantages with respect to the state of the art, a method according to the present invention for calculating the biological age of an individual comprises the use of the lipidomic analysis of the cell membrane, in particular of the cell membrane lipodome, meaning the fatty acid composition of the phospholipids that constitute a cell membrane.

[0018] In particular, the invention focuses on the membrane of a red blood cell (erythrocyte), which has an average lifespan of 4 months, and on its fatty acid composition, which has known population parameters. The fatty acid values are data that have been used to develop a machine-learning-based predictive model. Because age prediction is a regression problem, the current method is a supervised machine learning algorithm.

[0019] The method of the present invention has as a significant target, the cell membrane, in particular the erythrocyte membrane, which, like DNA, is a key element of the living cell: without a membrane, the cell does not exist, as without DNA. In addition, the membrane, which delimits and wraps all cellular content, is more exposed to factors that influence chronological age, such as nutrition, stress, radiation, physiological (sport, pregnancy) and pathological states, and lifestylesin general.

[0020] The use of data from red blood cell (RBC) membrane lipidomic analysis represents a novelty compared to known procedures and has the same feasibility as DNA, since the sample can be obtained via non-invasive procedures. A few drops of blood can be obtained by self-collection using approved medical devices. In accordance with embodiments, cell membrane lipidomics analysis of RBC provides, for each individual, data on the composition of a predetermined number of lipids, for example, at least four, preferably between five and twenty, more preferably between eight and thirteen. The number of lipids analyzed can be any of four, five, six, seven, eight, nine, ten, eleven, twelve, thirteen, fourteen, fifteen, sixteen, seventeen, eighteen, nineteen and twenty, obtaining their quantitative values in the sample of RBC membranes. Advantageously, new indices or parameters can be derived from the quantitative values of each fatty acid. The value of fatty acids is derived from analysis of the RBC membrane lipidome, isolated using known techniques, in particular following a protocol approved as compliant with the ISO 15189 regulation. It is not excluded the possibility of developing the age calculation using data of cell membranes from other tissues, such as skin lipidomics, which can be obtained with a “tape strip” (Br. J. Dermatol. 2021, 185, 26-35).

[0021] The novelty presented in the present application is that of a data-driven bloodbased method for predicting age that relies solely on data from lipidomic analyses, without the need for hematological, metabolic or DNA methylation-based parameters.

[0022] According to embodiments, the lipids are fatty acids, at least two of the saturated fatty acid group, the monounsaturated fatty acid group, the omega-6 polyunsaturated fatty acid group and the omega-3 polyunsaturated fatty acid group. Favorably, fatty acids include at least arachidonic acid and oleic acid, as these two acids are the ones most involved in inflammation and protection of the body.

[0023] In accordance with embodiments, the fatty acids belonging to RBC membrane phospholipids can be: palmitic acid, stearic acid, palmitoleic acid, oleic acid, vaccenic acid, di-homo-glinolenic acid, arachidonic acid, eicosapentaenoic acid (EP A), docosahexaenoic acid (DHA) and linoleic acid. To this basic group, some other fatty acids can be added according to specific cases and according to theevolution of scientific knowledge: as an example, docosapentaenoic acid DPA is mentioned, which signals the condition of obesity (Biomedicine 2024, 12, 88), and / or pentadecanoic acid (Cl 5:0) which is a fatty acid not metabolically prepared by the human organism, but highlighted as presence in long-lived populations. According to embodiments, the method comprises a step of calculation of a unique parameter, said step involving calculating the peroxidation index on the basis of the measured lipids.

[0024] According to embodiments, the method comprises a step of application of the predictive model, which includes a sub-phase for calculating an age vector based on the measured lipids and the calculated unique parameter. The calculation can be performed with programs available in open access that analyze the data coming from the lipidomic analysis, as described above. The result is the biological age, which integrates information about the subject's chronological age.

[0025] In accordance with embodiments, the method comprises a step of making available data from red blood cell membrane fatty acid analysis and transforming them, using artificial intelligence, into specific features that have positive and negative effects on chronological age.

[0026] Preferably, the predictive model comprises an ensemble of regression algorithms configured to generate the biological age by averaging their respective age predictions.

[0027] It is possible to provide that the predictive model applies a correction factor that increases the predicted biological age when the arachidonic acid value in the erythrocyte membrane is at least 17%. Preferably, the correction factor provides to add between 0.2 to 2 years to the calculated biological age for every 0.5% increase of the arachidonic acid value, more preferably 1 year for every 0.5% of increase. The number of years to be added to the biological age depends on the health conditions of the person, especially when the person for which biological age is calculated is unhealthy.

[0028] It is also possible to provide that the predictive model applies a correction factor that increases the predicted biological age when the peroxidation index value is at least 151. Preferably, the correction factor provides to add 0.2 to 1 year to the calculated biological age for every 10 units of peroxidation index above 151, more preferably 1 year for every 10 units of increase.In accordance with the above-mentioned purposes and to solve the above-mentioned technical problem in a new and original way, also obtaining significant advantages with respect to the prior art, a predictive model using machine learning methods (implementable in any programmation language, such as for example Python) is presented, which can be run on a computer, and which computes a blood age that can interpreted as a prediction of a subject’s biological age.

[0029] ILLUSTRATION OF DRAWINGS

[0030] These and other aspects, features and advantages of the present invention will become apparent from the following description of some embodiments provided by way of non-restrictive example, with reference to the accompanying drawings wherein:

[0031] - fig. 1 is a schematic representation of a procedure for calculating biological age, according to the present invention;

[0032] - figures 2, 3 A and 3B are graphical representations of the lipidomic composition of individuals as a function of their age;

[0033] - fig. 4 is a graphic representation of the evolution of the peroxidation index as a function of the age of the individuals.

[0034] - fig. 5 is a histogram representation of each of the input variables and the target variable;

[0035] - fig. 6 is a heat map representation of the correlation between the lipids examined and age.

[0036] It is to be highlighted that in the present description, phraseology and terminology used, as well as the figures of the attached drawings as they are described, have the only purpose of better illustrating and explaining the present invention having a function of no limiting exemplifying the same invention, the scope of protection being defined in the claims.

[0037] To facilitate understanding, the same reference numbers have been used, where possible, to identify identical common elements in the figures. It is understood that elements and features of one embodiment can conveniently be incorporated into other embodiments without further clarifications.

[0038] DESCRIPTION OF AN EMBODIMENT OF THE PRESENT INVENTION

[0039] With reference to Figure 1, a biological age calculation method according to the present invention, provides for the subsequent phases of making available:1) a certified analysis of the membrane lipids, in particular the fatty acids values, obtained quantitatively from mature RBC membrane phospholipids by an accredited ISO 15189 protocol;

[0040] The fatty acids include, but are not limited, to:

[0041] Palmitic acid (16:0)

[0042] Stearic acid (18:0)

[0043] Oleic acid (18:1)

[0044] Palmitoleic acid ( 16: 1 )

[0045] Vaccenic acid (18:1)

[0046] Linoleic acid omega-6 (18:2)

[0047] Di-homo-gamma-linolenic acid or eicosatrienoic acid (20:3) omega-6 Arachidonic acid omega-6 (20:4)

[0048] Eicosapentaenoic acid EPA omega-3 (20:5)

[0049] Docosahexaenoic acid DHA omega-3 (22:6)

[0050] Thus, each sample is characterized by ten different fatty acids, which are part of four different groups (in this case saturated fatty acids, monounsaturated fatty acids, polyunsaturated fatty acids omega-6 and polyunsaturated fatty acids omega-3).

[0051] It is to be noted that the data that enters the evaluation for BA preferably come from an accredited procedure for the isolation of the RBC membrane from a mature RBC, under the ISO 15189 regulation. This ensures that the data are certified and repeatable, with certified measurement error, eliminating bias introduced by the use of different analytical methods.

[0052] The mathematical operations by Ridge regression model and the age prediction algorithm described above are obtained for the first time and represent the innovation content of this invention.

[0053] The mathematical operations and functions, together with the data acquisition and database construction, can be implemented in Python, more preferably in Python 3.0 or Python 3.12, using scientific and machine learning libraries (for instance SciPy / Numpy / Pandas / Scikit-leam) to obtain the final data as number or any kind of graphical object able to express the chronological vs biological age of the individual.

[0054] The method and the sequence of steps transform the fatty acid values intofeatures which have positive or negative effects on the age of the individual, starting from the baseline value of 49.17.

[0055] The model for improving prediction accuracy ensures high performance in all scenarios. The ensemble combines 8 regression algorithms to predict age from fatty acid profiles. It uses simple averaging of predictions from models that represent fundamentally different algorithmic approaches: tree-based methods, linear regularised models, gradient boosting, and neural networks. The basis for using an ensemble model was the bias-variance decomposition and ensemble diversity theory, which demonstrate that combining predictions from several diverse, reasonably accurate models lowers the total prediction error. Furthermore, gradient boosting and random forests are resilient against multicollinearity since there is no matrix inversion involved in the computation and therefore the coefficients p do not become overly sensitive to noise. As tree-based methods, they use a local search at each node to find a split that maximises the information gain. This was necessary as the original features (see Table) exhibited high multicollinearity (computed variance inflation factors).

[0056] Table 1. Voting Regressor Complete Ensemble Model

[0057]

[0058] * learning rate

[0059] Table 2. Model Description

[0060]

[0061] Two constraints were inserted as correction factors:

[0062] • If a subject’s erythrocyte membrane arachidonic acid >= 17, the model is constrained to overestimate the subject’s age. It will do so proportionally, e.g. for every 0.5 percentage points above 17, it will add a year of biological age. For example, an arachidonic acid value of 18 will equate to 2 years added to the predicted age. This correction is implemented as arachidonic acid is well known as an inflammatory mediator (Wang et al., 2025).• If a subject’s peroxidation index >= 151, the model is similarly constrained to overestimate the subject’s age. It will do so proportionally, e.g. for 10 units above 151, it will add a year of biological age. For example, a peroxidation index value of 161 will equate to 1 year added to the predicted age. This adjustment reflects the observed role that PI has in animal longevity (Puca et al., 2020).

[0063] • Any further potential correction factors based on biochemical and metabolic principles will be implemented in the future.

[0064] • Even a subset of subjects with 2 or more measurements years apart would allow training on within-person lipid trajectories, and the difference between lipidomic measurements on a per subject basis can be incorporated as a predictive signal into the model.

[0065] The method has been tested on a database of 11,153 subjects (7672 females and 3476 males after discounting invalid entries) with composition data for 10 RBC membrane fatty acids, including mono-unsaturated, poly-unsaturated, and saturated fatty acids (SFA, MUFA, and PUFA). Table 3 below summarises the fatty acids whose composition in erythrocyte membranes has been analytically measured and taken as inputs.

[0066] Table 3

[0067]

[0068]

[0069] A composite parameter, the peroxidation index PI, is derived as a sum of certain of the above fatty acid percentages weighted by coefficients as such:

[0070] PI = (0.025 * Tot MUFA) + (Linoleic Acid %) + (2 * Eicosatrienoic Acid %)

[0071] + (4 * Arachidonic Acid ω6 %) + (6 * EPA ω3 %) + (8 * DHA ω3 %)

[0072] Tabel 4 below discloses a set of basic engineered features, derived from combinations of two or more of the fatty acids above:

[0073] Table 4

[0074]

[0075]

[0076] Table 5 below discloses a set of advanced engineered features derived from combinations of two or more of the fatty acids considered above:

[0077] Table 5

[0078]

[0079]

[0080] The biological motivation underlying the calculation of these features and their use in the predictive model is shown in table 6):

[0081] Table 6

[0082]

[0083]

[0084] The database was examined by exploratory data analysis and health segmentation analysis. The fatty acid trends were examined as a function of age and plotted using standard data aggregation and group operation tools available in the Pandas library such as pivot tables and cross-tabulations. Mean values for all fatty acids across all age points (1-105) were used as measures of central tendency to examine relationships and underlying trends.

[0085] The trends of the fatty acids are shown in Figure 2.

[0086] The Pearson coefficients among the fatty acid constituents of the RBC membrane phospholipids indicate strong negative to strong positive relationshipsbetween the variables.

[0087] The results are plotted in figure 6.

[0088] Training Dataset (Age and Gender Composition)

[0089] For building the training dataset, the erythrocyte membrane fatty acid data from 13,274 subjects were used, collected over the period 2021-2025. Age and BMI outliers were removed as follows:

[0090] • The model was trained on subjects aged 35-75

[0091] • The model was trained on subjects in the BMI range 15.5 - 40.5 Following outlier removal, the database consisted of 9,972 subjects.

[0092] Table 7

[0093]

[0094] For segmenting the health cohorts, the following criteria were used:

[0095] Healthy Individuals: Do not have any illness and have BMI in the range 18.5 - 24.9

[0096] Unhealthy Individuals: Have at least one illness and have BMI outside the range 18.5 -24.9

[0097] The Results can be evaluated by the following metrics to assess the performance of the regression model:

[0098] (1) MSE (Mean Squared Error):

[0099] MSE measures the average squared difference between the predicted values and the actual values. A lower value indicates better model performance.

[0100] (2) RMSE (Root Mean Squared Error):

[0101] RMSE is simply the square root of MSE and provides an aggregate measure of how far predictions are from the actual outcomes. RMSE is appealing because it is in the same units as the target variable.(3) MAE (Mean Absolute Error):

[0102] MAE calculates the average absolute differences between predicted and actual values. It is generally considered to be more interpretable than the MSE because it conveys the average error in the same units as the target variable.

[0103] (4) R2(Coefficient of Determination):

[0104] R2represents the proportion of variance in the target variable that can be explained by the predictor variables in the model. Mathematically, it is the square of the Pearson R-value (i.e. the correlation coefficient). The R2effectively measures how predictable the target variable (i.e. the dependent variable) is based on the features (i.e. the independent variables).

[0105] Finally, Correction factors can be implemented on certain data to better express their biological significance, such as, for example, not exhaustively described, the level of Arachidonic Acid > 17% and the absolute value of Peroxidation Index > 151. The enlargement of the dataset can require further correction factors which are comprises in the steps of the invention.

[0106] The mathematical formula derived for the Biological Age prediction calculation was determined originally using a Ridge regression model, and has the general form:

[0107] Age = Intercept + L (Coefficient x Feature) (eq. 1) with the Baseline (Intercept) -> Age = 53.89 years

[0108] The intercept is the predicted age when all features equal 0 (i.e. in standardized space). Since the features are standardized (p = 0, o = 1), a mean of 0 represents the population average.

[0109] The list of Features and their coefficient with the individual effect on the age (increasing or decreasing) is shown in Table 1.

[0110] Table 8 - Features & Calculated Coefficients from Ridge Regression

[0111]

[0112]

[0113] Hence the ridge equation derived to compute age using hematologic and physiological features is: wAge = 53.89

[0114] + 12.8595 x Oleic (18:1)

[0115] - 9.1534 x log(Peroxidation Index)

[0116] + 8.9415 x log(Cumul_PUFA_co6)

[0117] + 7.4926 x Bio Age Score

[0118] + 7.1940 x Oxidative Vulnerability

[0119] + 6.6579 x (Cumul_PUFA_co6)2

[0120] - 6.2086 x log(Cumul_MUFA)

[0121] - 5.3688 x Oleic_Linoleic_interaction

[0122] + 5.2842 x Linoleic co6 (18:2)

[0123] - 5.2130 x Peroxidation Index

[0124] + 4.9164 x Cumul MUFA normalised + 4.5796 x DHA co 3 (22:6)

[0125] - 4.5122 x UFA_total

[0126] + 4.4105 x co6:co3

[0127] - 4.2693 x Oxidative Balance

[0128] + 4.2062 x Inflammatory Index

[0129] - 3.8119 x Saturation Index

[0130] + 3.7824 x Healthy Ageing Index

[0131] - 2.5001 x [Oleic (18:1)]2

[0132] + 2.2173 x BMI

[0133] + 1.8422 x Cumul_SFA

[0134] + 1.6927 x EPA ω3 (20:5)

[0135] - 1.6632 x BMI_x_D6D

[0136] + 1.1659 x BioAge x BMI

[0137] + 0.9906 x anti_inflammatory_ratio

[0138] + 0.8595 x LA to ARA ratio

[0139] + 0.8233 x D6D_activity

[0140] + 0.8051 x log(omega_ratio)

[0141] - 0.7456 x omega balance composite - 0.7141 x Double Bond Index

[0142] - 0.6392 x Cumul_PUFA_ω3_normalised - 0.5215 x O9:O6- 0.3680 x (Cumul MUFA)3

[0143] + 0.2141 × Cumul_PUFA_ω3

[0144] - 0.0994 x Cardio Risk Index

[0145] Interpretation of the Coefficients

[0146] Positive Coefficients (Blue Background)

[0147] • Higher values of these features increase the predicted biological age.

[0148] • For every 1 standard deviation (σ) increase in the feature, the coefficient value (in years) is added to the predicted age.

[0149] Negative Coefficients (Orange Background)

[0150] • Higher values of these features decrease the predicted biological age.

[0151] • For every 1 standard deviation (σ) increase in the feature, the absolute coefficient value (in years) is subtracted from the predicted age.

[0152] Important Notes

[0153] 1. All features must be standardized before applying the equation (subtract μ, divide by σ).

[0154] 2. Coefficients represent the change in predicted age per 1 standard deviation change in the feature.

[0155] 3. This is a linear approximation - tree-based models (Extra Trees) provide the best accuracy.

[0156] 4. The equation was derived using ridge regression with regularization parameter α = 1.0.

[0157] 5. The most scientifically valid way of representing the model’s predictive capacity is using this Ridge model, since it performed nearly as well as the ensembles. However, this equation represents the linear trend of biological ageing; however, the extremely randomized tree ensemble captured an additional 1-2% of accuracy by accounting for non-linear interactions that cannot be expressed as a linear equation.

[0158] Feature Engineering

[0159] Given the numeric representation of data fed into a model, obtaining the correct features is crucial for designing effective predictive machine learning models. Feature engineering is the process of formulating the most appropriate features given the data, the model and the scientific objective. The number of features is also important, as if there are not enough informative features the model will beunable to perform the ultimate task. If there are too many features, or if most of them are irrelevant, then training the model will be more difficult.

[0160] In order to train the age prediction model using the erythrocyte membrane-based biomarkers, new features were designed and added to the original ones (table 1). Feature engineering was performed in two steps: (1) creation of basic aggregate and ratio features based on saturated, mono- and poly -unsaturated fatty acid groups as well as physiological indices (BMI) and (2) design of advanced features to extract meaningful biological signals from fatty acid data. A total of 58 features consisting of the 11 original and 47 designed (basic and advanced) were used to train the model. The 35 strongest features, as computed by an averaging of three normalised rankings (ANOVA, F-statistic, Pearson correlation), were selected as part of the final engineering set and used in the predictive model.

[0161] Feature engineering expands the 11 raw inputs into 58 derived features: ratios between fatty acid families (omega-6 / omega-3, oleic / linoleic), oxidative stress indices, inflammatory composites, log transforms to correct skewed distributions, polynomial terms (squared, cube root), and a PCA-derived Bio Age Score that compresses the inflammatory marker cluster into a single dimension.

[0162] Model Training

[0163] 10-fold stratified cross-validation is the core evaluation framework (test size = 0.1, training size = 0.9). The dataset is split into 10 equal parts. In each iteration, 9 parts train the model and the remaining part tests it. This is repeated 10 times so every patient is tested exactly once. Age stratification ensures each fold has a balanced age distribution rather than accidentally concentrating young or old patients in a single fold. The reported MAE and R2are averages across all 10 folds and represent the out-of-sample performance on the training set.

[0164] Feature selection runs inside each fold on the training partition only, using a combined score of ANOVA F-statistic, Pearson correlation, and standardised regression coefficients. The top 35 features from each fold are recorded. After all 10 folds, a consensus set is built by majority vote — only features selected in more than 5 of 10 folds enter the final model. This prevents the selection process from seeing test data, which would inflate performance estimates.

[0165] Eight candidate models compete in the CV: Extra Trees (Extremely Randomised Trees), Random Forest, Gradient Boosting, Ridge Regression, ElasticNet, XGBoost, LightGBM, and a Neural Network. Each is evaluated independently on the same folds and the winner by overall MAE and R2is selected. The VotingRegressor is the final deployed model. Rather than using the single best model, the top 4 models by CV performance are combined into an ensemble that averages their predictions. This reduces variance — when one model is wrong in one direction, the others tend to partially compensate. The VotingRegressor is fitted on the full training set using the consensus feature set and evaluated on a held-out 10% test split.

[0166] The Ridge age equation is a separate, interpretable output. Because tree ensembles cannot produce coefficients, Ridge Regression is fitted independently on the same consensus features and its standardised coefficients are extracted. This gives a linear equation of the form Predicted Age = intercept + β₁×Feature₁ + β₂×Feature₂ + … which clinicians and regulators can read, audit, and implement in a spreadsheet — even though the VotingRegressor is used for actual predictions.

[0167] Model Performance

[0168] Cross-Validation Performance (table 9)

[0169]

[0170] Performance on Unseen Labelled Data (Extra Trees, table 10)

[0171]

[0172] The interpretation of these results is that, on average, the model’s age predictions are off by about 12.4 years on labelled, unseen subjects.

[0173] The novelty introduced in the model’s calculation is based on two factors: (1) its focus on the erythrocyte membrane lipid composition and (2) its scope as a tool for predicting biological age based on molecular lipidomic markers.

[0174] 1) The mean absolute error on labelled, unseen data (MAE) is 12.39 years, i.e. comparable with the procedures for predicting biological age using DNA, which in the state of the art is the most interesting target for such calculation. That is, the method according to the present invention identifies subjects with a membrane condition that does not match the average of the deviations detected by a healthypopulation of that age, and moves to higher ages (case of higher biological age), as reported in table 11:

[0175] Table 11

[0176]

[0177] It may also be the case that the chronological age is almost equal (MAE <= 1 year) to the biological age, that is, the method according to the present invention identifies a coherence between the membrane state found in the individual and that found in the reference population (identical biological age case), as reported in Table 12:

[0178] Table 12

[0179]

[0180] It is clear that modifications and / or additions of parts or phases may be made tothe method for calculating biological age as described heretofore, without departing from the scope of the invention as defined in the claims.

[0181] It is also clear that, even though the present invention is described with reference to some specific examples, an expert of the field can realize other equivalent forms of method for managing a parking bay of a charging station and corresponding charging station, having the features expressed in the claims and so being all included in the scope of protection defined thereby.

[0182] In the following claims, the references in brackets have the only scope of facilitating reading and shall not be considered as limitative factors of the scope of protection defined in the same claims.

Claims

CLAIMS1. A method for calculating the biological age of a person, characterized in that it involves the acquisition of the fatty acid composition of the erythrocyte membrane lipidome of said person following the ISO 15189 method, and the use of a predictive model based on artificial intelligence.

2. Method as in claim 1, characterized in that said analysis is performed on a predetermined number of fatty acids from red blood cell membrane phospholipids which is equal to at least four, preferably between five and twenty.

3. Method as in claim 2, characterized in that said lipids comprise at least arachidonic acid and oleic acid.

4. Method as in claim 2 or 3, characterized in that said lipids are selected from several known lipid groups, preferably from at least two of the group of saturated fatty acids, the group of monounsaturated fatty acids, the group of polyunsaturated fatty acids ω6 and the group of polyunsaturated fatty acids ω3.

5. Method as claimed in any of claims 2 to 4, characterized in that said lipids are selected from palmitic acid, stearic acid, palmitoleic acid, oleic acid, vaccenic acid, di-homo- gamma-linolenic acid, arachidonic acid ω6, eicosapentaenoic acid EPA ω3, docosahexaenoic acid DHA ω3 and linoleic acid ω6.

6. Method as in claim 5, characterized in that the lipids may also contain docosapentaenoic acid DPA and / or pentadecanoic acid.

7. Method as in any of claims 2 to 6, characterized in that it comprises a step of calculation of a unique parameter derivable from the fatty acid composition.

8. Method as in claim 6, characterized in that said unique parameter is the peroxidation index, which is calculated on the basis of the measured lipids.

9. Method as in any of claims 2 to 8, characterized in that it comprises a step of application of the predictive model, which includes a sub-phase of calculation of an age vector on the basis of the measured lipids and the calculated unique parameter.

10. Method as in any of claim 2 to 9, characterized in that it comprises a step of making available the data of red blood cell membrane fatty acid analysis and transform them by artificial intelligence into specific features and related coefficients having positive and negative effects on the chronological age.

11. Method as in any of claims 1 to 10, characterized in that the predictive modelcomprises a supervised machine learning regression algorithm.

12. Method as in any of claims 1 to 11, characterized in that the predictive model comprises an ensemble of a plurality of regression algorithms configured to generate the biological age by averaging respective age predictions of the plurality of regression algorithms.

13. Method as in any of claims 1 to 12, characterized in that the predictive model applies a correction factor that increases the predicted biological age when the arachidonic acid value in the erythrocyte membrane is at least 17%.

14. Method as in claim 13, characterized in that the correction factor provides to add between 0.2 to 2 years to the calculated biological age for every 0.5% increase of the arachidonic acid value, preferably 1 year for every 0.5% of increase.

15. Method as in any of claims 8 to 14, characterized in that the predictive model applies a correction factor that increases the predicted biological age when the peroxidation index value is at least 151.

16. Method as in claim 15, characterized in that the correction factor provides to add 0.2-1 year to the calculated biological age for every 10 units of peroxidation index above 151, preferably 1 year for every 10 units of increase.

17. A computer program containing instructions that, when executed by a computer, result in the execution of a calculation step of a unique lipid-related parameter measured in the membrane of red blood cells taken from a sample of a person's body fluid, and / or a step of applying an artificial intelligence-based predictive model.

Images