Biological age prediction correction method and system based on individual difference

Abstract

The invention provides a biological age prediction correction method and system based on individual differences, relates to the field of artificial intelligence, and solves the technical problems of systematic prediction deviation, missing of individual difference modeling and insufficient clinical interpretability caused by the fact that a biological age prediction model only depends on a time age label and a statistical loss function in the prior art. The method comprises the following steps: acquiring multi-modal physiological data and corresponding time and age of a plurality of healthy individual samples; predicting an initial biological age prediction value of the individual sample based on the physiological data; and based on the time age of the individual sample and the corresponding initial biological age prediction value, fitting to obtain an average biological age prediction function, calculating an aging difference value of the individual sample relative to the same-age group, further constructing a composite loss function, and training the biological age prediction model by minimizing the composite loss function to obtain a biological age prediction result. And obtaining a biological age prediction correction model. The method is used in the biological age prediction correction process.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, and in particular to a biological age prediction and correction method and system based on individual differences. Background Technology

[0002] Current methods for predicting biological age primarily employ regression models, using chronological age as a label and training them using statistical loss functions such as minimizing mean squared error (MSE). However, these methods are generally plagued by systematic prediction bias: they tend to overestimate biological age in younger populations and underestimate it in older populations, leading to inconsistent model performance across different age groups or populations with significant differences in individual characteristics. While existing solutions attempt to mitigate this bias by introducing skewed loss functions, they are based solely on residual direction weighting and lack modeling of the relationship between individual physiological characteristics and prediction error. This results in unexplained sources of bias, rigid adjustment mechanisms, and an inability to meet clinical requirements for fairness and traceability. Therefore, this application provides a biological age prediction correction method and system based on individual differences. Summary of the Invention

[0003] This application provides a biological age prediction correction method and system based on individual differences, which solves the technical problems of systematic prediction bias, lack of individual difference modeling, and insufficient clinical interpretability caused by existing biological age prediction models that rely solely on time age labels and statistical loss functions.

[0004] To achieve the above objectives, this application adopts the following technical solution: Firstly, it provides biological age prediction correction methods based on individual differences, including: Multimodal physiological data and corresponding time ages of multiple healthy individual samples are obtained. The physiological data are biological data that objectively reflect the physiological function status of the human body, including at least one biomarker related to arteriosclerosis. The time age is the actual number of years that the individual sample has experienced since birth. Physiological data is preprocessed to obtain biofeature vectors, which are then input into a biological age prediction model to output an initial biological age prediction value for an individual sample. The biological age prediction value is the age of the actual biological aging degree, which may be higher or lower than the chronological age. The biological age prediction model is trained based on an artificial intelligence neural network model. Based on the time age of individual samples and the corresponding initial biological age prediction value, an average biological age prediction function is obtained by fitting, which represents the average biological age prediction value of the same age group at any time age. Based on the predicted biological age value and the average biological age prediction function, the aging difference value of an individual sample relative to the same age group is calculated. Based on the aging difference value, a composite loss function is constructed. The biological age prediction model is trained with minimizing the composite loss function as the optimization objective to obtain a bias-corrected biological age prediction model. The standardized biofeature vector of the individual to be tested is input into the biological age prediction and correction model, and the bias-corrected biological age prediction value is output.

[0005] Based on the above technical solutions, this invention proposes a two-stage physiological age prediction bias correction method in the individual-difference-based biological age prediction correction method provided in this application. Its core advantage lies in introducing individual risk difference (IRD) into the model training mechanism, significantly improving the clinical interpretability and practicality of the prediction results. Unlike traditional methods that directly use chronological age as a label, this approach first obtains the average biological age deviation value (i.e., IRD) of each individual relative to healthy individuals of the same age through pre-training, and then incorporates it into the composite loss function during the retraining stage. This allows the model to balance overall fitting accuracy and sensitivity to abnormal aging during optimization. This design effectively alleviates the systematic bias caused by using chronological age as a proxy label (such as overestimating young people and underestimating the elderly), and more realistically reflects the individual's physiological state.

[0006] In conjunction with the first aspect above, in one possible implementation, calculating the aging difference value of an individual relative to their age group includes: Based on the average biological age prediction function, the predicted average biological age value corresponding to the time age is obtained; through the formula... The aging difference value of individual sample i relative to the same age group was calculated. Where i is the sequence number of the individual sample, Let be the predicted biological age of the i-th individual sample. Let be the predicted average biological age of the i-th individual sample.

[0007] In conjunction with the first aspect above, in one possible implementation, constructing a composite loss function based on the aging difference value includes: By incorporating the aging difference value into the traditional MSE loss function, a composite loss function is obtained. : ; in, Here, represents the optimal weighting coefficient, and n represents the total number of individual samples. Let be the age of the i-th individual sample over time. Let be the predicted biological age of the i-th individual sample.

[0008] In conjunction with the first aspect above, in one possible implementation, the optimal hyperparameters... The methods for determining this include: The hyperparameter search range is preset, and multiple candidate values ​​are selected within the preset search range; For each candidate value, a corresponding biological age prediction model is trained using the training data, and the statistical performance index and clinical performance index of the model on the validation set are calculated respectively. The statistical performance index includes at least one of mean absolute error (MAE), mean squared error (MSE), root mean squared error (RMSE), and coefficient of determination (R²), and the clinical performance index includes at least one of recall, precision, F1 score, and area under the feature curve (AUC) of individual samples. Based on the preset clinical application goals, different weights are assigned to each performance indicator, and the weighted sum of the performance indicators corresponding to each candidate value is calculated to obtain a comprehensive score; among them, the candidate value corresponding to the highest comprehensive score is the optimal hyperparameter.

[0009] In conjunction with the first aspect above, in one possible implementation, the construction of the feature curve of the individual sample includes: Analyze the health risk score of each individual sample, and use the health risk score of each individual sample as a classification threshold in turn. For each classification threshold, predict the label category for each individual sample; where individual samples whose classification threshold is greater than the preset classification threshold are labeled as high risk, and the remaining individual samples are labeled as low risk. Based on the predicted label category and the actual label category for each individual, the number of true positive cases (TP), false positive cases (FP), true negative cases (TN), and false negative cases (FN) are calculated. Calculate the true positive rate (TPR) = TP / (TP + FN) and the false positive rate (FPR) = FP / (FP + TN) for each individual sample; construct the characteristic curve of the individual sample with FPR as the x-axis and TPR as the y-axis.

[0010] In conjunction with the first aspect above, in one possible implementation, the true positive count TP is the number of individual samples whose actual and predicted label categories are both high-risk; the false positive count FP is the number of individual samples whose actual label category is low-risk and whose predicted label category is high-risk; the true negative count TN is the number of individual samples whose actual and predicted label categories are both low-risk; and the false negative count FN is the number of individual samples whose actual label category is high-risk and whose predicted label category is low-risk.

[0011] In conjunction with the first aspect above, in one possible implementation, the analysis of the health risk score for each individual sample includes: Through formula The health risk score for each individual sample was calculated. ;in, For unit difference fractions, when ,but The value is a positive integer greater than 0. ,but The value is a positive integer less than 0.

[0012] In conjunction with the first aspect above, in one possible implementation, the method for obtaining the average biological age prediction function includes: Individual samples are divided into several intervals based on their time age. Within each interval, the arithmetic mean of the initial biological age predictions of all samples is calculated to obtain the average biological age prediction value for the corresponding interval. The median of the time age in each interval and the corresponding average biological age prediction value constitute several discrete points. Arbitrary time ages are inserted between each discrete point, and the corresponding average biological age prediction value is calculated through linear interpolation. The discrete points are fitted to obtain a continuous average biological age prediction function. The width of each interval is less than a preset width.

[0013] In conjunction with the first aspect above, in one possible implementation, the preprocessing of physiological data to obtain a biometric vector includes: Missing values ​​were processed in the multimodal physiological data collected from individual sample i; Biometrics with absolute Pearson correlation coefficients greater than a preset correlation threshold were selected as valid features. For each valid biometric, its mean and standard deviation were calculated, and then analyzed using the formula... The standardized value of the j-th biometric item for the i-th individual sample was calculated. j represents the sequence number of the biomarker; The standardized biological indicators are constructed into biological feature vectors in a preset order to obtain the biological feature vector of the i-th individual.

[0014] Secondly, this application provides a biological age prediction and correction system based on individual differences, comprising: a prediction module, and an acquisition module and a correction module connected thereto; wherein, the acquisition module is used to acquire multimodal physiological data and corresponding time ages of multiple healthy individual samples; the prediction module is used to preprocess the physiological data to obtain biofeature vectors, and input them into a biological age prediction model, outputting an initial biological age prediction value for the individual sample; based on the time age of the individual sample and the corresponding initial biological age prediction value, an average biological age prediction function is fitted; the correction module is used to calculate the aging difference value of the individual sample relative to the same age group based on the biological age prediction value and the average biological age prediction function; based on the aging difference value, a composite loss function is constructed, and the biological age prediction model is trained with minimizing the composite loss function as the optimization objective to obtain a bias-corrected biological age prediction and correction model; the standardized biofeature vector of the individual to be tested is input into the biological age prediction and correction model, and the bias-corrected biological age prediction value is output.

[0015] This application provides a method and system for correcting biological age prediction based on individual differences, which effectively solves the systematic prediction bias problem caused by traditional models that directly label age. First, the individual risk difference (IRD), i.e., the degree of deviation of an individual's biological age from the average level of healthy individuals of the same age, is estimated through a pre-trained model. Then, during the retraining phase, this IRD is explicitly incorporated into the loss function, guiding the model to simultaneously focus on overall fitting accuracy and the identification of individual aging anomalies during optimization. This two-stage design not only preserves the model's ability to learn general aging trends but also enhances its sensitivity to high-risk individuals, effectively mitigating the systematic bias problem of overestimating younger populations and underestimating older populations. Furthermore, this approach abandons the single statistical indicator evaluation mode and adopts a multi-dimensional evaluation strategy encompassing statistical perspectives such as MAE and R², as well as clinical perspectives such as recall and AUC, ensuring that the model optimization direction is consistent with actual medical needs. This method does not rely on a specific algorithm, has good versatility and scalability, and provides an innovative and feasible technical path for constructing a biological age prediction system that combines scientific rigor and clinical value. Attached Figure Description

[0016] Figure 1 A system architecture diagram of a biological age prediction and correction system based on individual differences is provided for embodiments of this application; Figure 2 A flowchart illustrating a biological age prediction and correction method based on individual differences, provided for an embodiment of this application; Figure 3 A schematic diagram illustrating the prediction results of various models based on a traditional loss function, provided for an embodiment of this application; Figure 4This is a schematic diagram illustrating the predictive performance evaluation of various models under a new composite loss function based on IRD, as provided in an embodiment of this application. Detailed Implementation

[0017] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] The biological age prediction and correction method based on individual differences provided in this application can be applied to biological age prediction and correction systems based on individual differences, such as... Figure 1 As shown, the system includes: a prediction module, and an acquisition module and a correction module connected thereto; The acquisition module is used to acquire multimodal physiological data and corresponding time ages of multiple healthy individual samples; The prediction module is used to preprocess physiological data to obtain biofeature vectors, which are then input into the biological age prediction model to output the initial biological age prediction value of individual samples. Based on the time age of individual samples and the corresponding initial biological age prediction value, the average biological age prediction function is fitted to obtain the mean biological age prediction function. The correction module is used to calculate the aging difference value of an individual sample relative to the same age group based on the predicted biological age value and the average biological age prediction function. Based on the aging difference value, a composite loss function is constructed. The biological age prediction model is trained with minimizing the composite loss function as the optimization objective to obtain a bias-corrected biological age prediction model. The standardized biofeature vector of the individual to be tested is input into the biological age prediction and correction model, and the bias-corrected biological age prediction value is output.

[0019] To address the technical problems of systematic prediction bias, lack of individual difference modeling, and insufficient clinical interpretability caused by existing biological age prediction models relying solely on time age labels and statistical loss functions, this application provides a biological age prediction correction method based on individual differences. This method includes: Multimodal physiological data and corresponding time ages of multiple healthy individual samples are obtained. The physiological data are biological data that objectively reflect the physiological function status of the human body, including at least one biomarker related to arteriosclerosis. The time age is the actual number of years that the individual sample has experienced since birth. Physiological data is preprocessed to obtain biofeature vectors, which are then input into a biological age prediction model to output an initial biological age prediction value for an individual sample. The biological age prediction value is the age of the actual biological aging degree, which may be higher or lower than the chronological age. The biological age prediction model is trained based on an artificial intelligence neural network model. Based on the time age of individual samples and the corresponding initial biological age prediction value, an average biological age prediction function is obtained by fitting, which represents the average biological age prediction value of the same age group at any time age. Based on the predicted biological age value and the average biological age prediction function, the aging difference value of an individual sample relative to the same age group is calculated. Based on the aging difference value, a composite loss function is constructed. The biological age prediction model is trained with minimizing the composite loss function as the optimization objective to obtain a bias-corrected biological age prediction model. The standardized biofeature vector of the individual to be tested is input into the biological age prediction and correction model, and the bias-corrected biological age prediction value is output.

[0020] Based on this, the individual-difference-explainable loss function proposed in this application significantly reduces prediction bias, especially across different age groups or health conditions, by introducing an individual-difference correction term into the loss function. Furthermore, the loss function in this application not only optimizes prediction accuracy but also provides interpretability, enabling the model to clearly reveal the sources of prediction bias. This approach not only improves prediction accuracy but also provides actionable feedback for clinical practice, demonstrating broader application potential.

[0021] like Figure 2 As shown in the embodiments of this application, the biological age prediction and correction method based on individual differences includes: S201. Obtain multimodal physiological data and corresponding time ages of multiple healthy individual samples.

[0022] The multimodal physiological data are derived from at least one of the following: arteriosclerosis detection devices, electrocardiograms, medical imaging, blood biochemical indicators, or wearable sensors. The physiological data includes at least one arteriosclerosis-related biomarker, such as at least one of the following: left and right arm ankle pulse wave velocities, ankle-brachial index, ankle systolic blood pressure, or mean arterial pressure.

[0023] Time age refers to the actual number of years an individual sample has lived since birth.

[0024] S202. Preprocess the physiological data to obtain biological feature vectors, and input them into the biological age prediction model to output the initial biological age prediction value of the individual sample.

[0025] Among them, the biological age prediction value is the age of the actual biological aging degree, which is higher or lower than the time age.

[0026] The biological age prediction model is trained based on an artificial intelligence neural network model, as detailed below: Acquire time-related and multimodal physiological data from a large number of healthy individuals, such as arteriosclerosis-related indicators: systolic blood pressure, mean arterial pressure, etc. The raw physiological data is preprocessed, including missing value imputation, outlier removal, feature selection, and standardization, to form a unified-dimensional biofeature vector; age is used as a supervision label. The architecture uses a multi-layer fully connected neural network, which includes an input layer (with dimension d, consistent with the dimension of the biological feature vector); several hidden layers (e.g., 3–5 layers, each containing 64–256 neurons); each hidden layer is followed by a non-linear activation function (such as ReLU or Swish); and an output layer (single neuron, with no activation function or using linear activation) that directly outputs continuous values ​​as the biological age prediction result. The mean squared error is used as the training loss function for the model. Where n is the total number of individual samples, Let be the age of the i-th individual sample over time. Let be the predicted biological age of the i-th individual sample; After the model is trained on a training set of healthy people, the model is used as the first-stage pre-trained model, namely the biological age prediction model. For any input individual's biometric vector, the model outputs a numerical value, which is the initial predicted biological age. This value reflects an individual's current biological aging state, and theoretically it can be higher or lower than their age over time.

[0027] S203. Based on the time age of individual samples and the corresponding initial biological age prediction value, the average biological age prediction function is fitted to obtain the average biological age prediction function. The age prediction function represents the predicted average biological age of a group of people of the same age at any given time.

[0028] Specifically, individual samples are divided into several intervals based on their time age. Within each interval, the arithmetic mean of the initial biological age predictions of all samples is calculated to obtain the average biological age prediction value for the corresponding interval. The median of the time age in each interval and the corresponding average biological age prediction value constitute several discrete points. Arbitrary time ages are inserted between each discrete point, and the corresponding average biological age prediction value is calculated through linear interpolation. The discrete points are then fitted to obtain a continuous average biological age prediction function. The width of each interval is less than a preset width.

[0029] For example, all healthy individuals in the training set are divided into several continuous and non-overlapping intervals according to their age at time, and the width of each interval does not exceed a preset maximum width, such as 2 years, 3 years, or 5 years.

[0030] For each interval, collect individual samples that fall within that interval, extract their predicted biological age values, and calculate the average of several predicted biological age values. Using the median of the time age in each interval as the x-coordinate of the coordinate system and the corresponding predicted biological age as the y-coordinate, a series of discrete coordinate points are obtained. Furthermore, by inserting any given time age between the centers of two adjacent intervals, the corresponding average predicted biological age is calculated through linear interpolation, thus obtaining a continuous average biological age prediction function.

[0031] S204. Based on the predicted biological age value and the average biological age prediction function, calculate the aging difference value of the individual sample relative to the same age group; Based on the aging difference value, a composite loss function is constructed. The biological age prediction model is trained with minimizing the composite loss function as the optimization objective to obtain a bias-corrected biological age prediction model. The standardized biofeature vector of the individual to be tested is input into the biological age prediction and correction model, and the bias-corrected biological age prediction value is output.

[0032] Based on the above technical solutions, this application proposes a biological age prediction bias correction method based on Individual Risk Difference (IRD), effectively overcoming the systematic bias problem caused by traditional models relying on chronological age as a proxy label. This method deconstructs biological age into two interpretable components: chronological age and individual aging deviation. It constructs an average physiological age prediction function to characterize the typical aging trajectory of healthy individuals of the same age, and uses this as a basis to quantify the degree of abnormality (i.e., IRD) of each individual relative to the group. A composite loss function designed on this basis allows the model to focus not only on overall fitting accuracy during training but also to enhance its ability to identify high-risk individuals. Compared to existing methods, this technique can significantly improve clinically relevant performance without changing the model structure, and is particularly suitable for modeling multimodal physiological indicators closely related to aging, such as arteriosclerosis. Furthermore, this framework has good transferability and is expected to be extended to other physiological age assessment scenarios such as brain age and heart age. More importantly, this method emphasizes a multi-dimensional assessment concept that gives equal weight to statistical and clinical indicators, avoiding model bias caused by a single optimization objective, and providing a reliable technical path and theoretical support for the transition of biological age models from scientific research to clinical application.

[0033] In one possible implementation of this application embodiment, the above-mentioned S202 can be specifically implemented by the following S301, S302 and S303, which are described in detail below: Biometrics with absolute Pearson correlation coefficients greater than a preset correlation threshold were selected as valid features. For each valid biometric, its mean and standard deviation were calculated, and then analyzed using the formula... The standardized value of the j-th biometric item for the i-th individual sample was calculated. j represents the biometric index number; The standardized biological indicators are constructed into biological feature vectors in a preset order to obtain the biological feature vector of the i-th individual.

[0034] Specifically, for each raw biometric, such as pulse wave velocity and blood pressure, the Pearson correlation coefficient between the individual's time-age and the training set of healthy individuals is calculated: ; in, For covariance, Let j be the standard deviation of biological indicator j. denoted as the standard deviation of the age over time.

[0035] like If the value exceeds the preset threshold, the biometric j is significantly correlated with the aging process and is retained as a valid feature; otherwise, it is discarded.

[0036] Based on the above technical solution, by selecting bioindicators whose absolute values ​​of the Pearson correlation coefficient with time-age are greater than a preset threshold as effective features, the method achieves accurate extraction of biological signals significantly related to aging, effectively eliminating noisy indicators that are irrelevant to or weakly correlated with the aging process, and improving the quality and biological interpretability of the model input data. This method is based on statistical principles, using a training set of healthy individuals to calculate the correlation between each bioindicator and time-age, ensuring that the selected features have clear biological significance. For example, indicators known to be closely related to arteriosclerosis and the aging process, such as pulse wave velocity and blood pressure, are retained. Simultaneously, the retained bioindicators are standardized by subtracting the mean and dividing by the standard deviation, eliminating the influence of differences in dimensions and scales between different indicators. This allows multimodal physiological data to be modeled in a unified space, enhancing the convergence and generalization ability of the neural network model. Furthermore, this feature selection strategy does not rely on complex machine learning algorithms, is computationally efficient and easy to implement, and is suitable for large-scale clinical data processing scenarios, laying a solid data foundation for the subsequent construction of high-precision, low-bias biological age prediction models.

[0037] In one possible implementation of this application embodiment, the above-mentioned S204 can be specifically implemented by the following S301, S302 and S303, which are described in detail below: S301. Based on the average biological age prediction function, obtain the average biological age prediction value corresponding to the time age; through the formula... The aging difference value of individual sample i relative to the same age group was calculated. Where i is the sequence number of the individual sample, Let be the predicted biological age of the i-th individual sample. Let be the predicted average biological age of the i-th individual sample.

[0038] S302. Construct a composite loss function based on aging difference values. : ; in, Here, represents the optimal weighting coefficient, and n represents the total number of individual samples. Let be the age of the i-th individual sample over time. Let be the predicted biological age of the i-th individual sample.

[0039] It should be pointed out that when hour, Equivalent to traditional MSE loss; while when hour, It relies entirely on IRD correction.

[0040] Based on the above technical solution, this application overcomes the limitation of traditional models that only use chronological age as a supervisory signal. By deconstructing biological age into two components—chronological age and individual aging deviation—this method can explicitly model the aging heterogeneity within a peer group, thus more realistically reflecting the individual's biological state. The constructed composite loss function integrates statistical fitting error and individual deviation information relative to the peer group, enabling the model to not only pursue overall prediction accuracy during training but also focus on sensitivity to abnormal aging patterns. This design naturally aligns with clinical needs, prioritizing the identification of high-risk individuals who significantly deviate from the healthy trajectory, rather than simply minimizing the average error. Furthermore, this method is not dependent on a specific model architecture and can be flexibly adapted to various trainable models such as multiple linear regression, gradient boosting trees, or neural networks, exhibiting good versatility and scalability. In addition, by introducing an average biological age prediction function as a benchmark, it effectively mitigates the systematic bias caused by using chronological age as a proxy label, such as overestimation of younger populations and underestimation of older populations, providing a theoretical basis for achieving a fairer and more individualized aging assessment.

[0041] In one possible implementation of this application embodiment, the method for determining the optimal weight coefficient in S302 above can be specifically implemented through the following S401, S402 and S403, which are described in detail below: S401. Preset hyperparameter search range, select multiple candidate values ​​within the preset search range; S402. For each candidate value, train the corresponding biological age prediction model using the training data, and calculate the statistical performance index and clinical performance index of the model on the validation set respectively. The statistical performance indicators include mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), and coefficient of determination. At least one of the following clinical performance metrics is required: recall, accuracy, F1 score, and area under the characteristic curve (AUC) for individual samples. S403. Based on the preset clinical application goals, assign different weights to each performance indicator, and calculate the weighted sum of the performance indicators corresponding to each candidate value to obtain a comprehensive score; among them, the candidate value corresponding to the highest comprehensive score is the optimal hyperparameter.

[0042] Based on the above technical solution, this hyperparameter optimization method significantly improves the practical value of biological age prediction models by systematically integrating statistical performance and clinical utility. Traditional hyperparameter tuning typically relies solely on statistical indicators such as MAE and R², which can easily lead to models that, while fitting age well, fail to adequately identify high-risk individuals (such as "unhealthy" populations). In contrast, this solution, in steps S401–S403, not only constructs a hyperparameter optimization model encompassing MAE, MSE, RMSE, and other parameters, but also... In addition to the statistical evaluation dimensions of classic regression indicators, this paper innovatively introduces key clinical indicators such as recall, precision, F1 score, and AUC. These indicators are dynamically weighted according to specific application scenarios (e.g., disease screening emphasizes high recall, while health management emphasizes balanced precision), and a comprehensive score is generated through weighted summation. This multi-objective collaborative optimization mechanism shifts the focus of hyperparameter selection from "optimal data fit" to "optimal clinical decision-making," effectively addressing the pain point of physiological age models being "statistically effective but clinically ineffective."

[0043] In one possible implementation of this application embodiment, the method for constructing the feature curve of the individual sample in S402 above can be specifically implemented through the following S501, S502 and S503, which are described in detail below: S501. Analyze the health risk score of each individual sample, and use the health risk score of each individual sample as a classification threshold in turn. Specifically, through the formula The health risk score for each individual sample was calculated. ;in, For unit difference fractions, when ,but The value is a positive integer greater than 0. ,but The value is a positive integer less than 0.

[0044] S502. For each classification threshold, predict the label category of each individual sample; wherein, the individual samples whose classification threshold is greater than the preset classification threshold are labeled as high risk, and the remaining individual samples are labeled as low risk. S503. Based on the predicted label category and the actual label category for each individual, count the number of true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN). True positives (TP) are the number of individuals whose true and predicted label categories are both high risk; false positives (FP) are the number of individuals whose true label category is low risk and whose predicted label category is high risk; true negatives (TN) are the number of individuals whose true and predicted label categories are both low risk; and false negatives (FN) are the number of individuals whose true label category is high risk and whose predicted label category is low risk.

[0045] Calculate the true positive rate (TPR) = TP / (TP + FN) and the false positive rate (FPR) = FP / (FP + TN) for each individual sample; construct the characteristic curve of the individual sample with FPR as the x-axis and TPR as the y-axis.

[0046] For example, 1. Traditional research typically does not consider IRD. In this embodiment, three commonly used algorithms, such as Multiple Linear Regression (MLR), Random Forest (RF), and XGBoost, are used to build a predictive model between five screening features and chronological age, serving as a tool for estimating biological age. The publicly available tool KDM is also included as a reference and comparison. Figure 3 The training results are shown; where (a) to (f) are the results of the three models under the traditional loss function, which is: (g) to (h) represent the prediction results of the KDM model.

[0047] The results show that KDM performs well, with the lowest correlation coefficient being 0.83. It is worth noting that, on the same dataset, while the performance of the three models is not as good as KDM, it is still close to the levels reported in many previous studies.

[0048] 2. Based on the IRD of this application, the above three models are then used to construct a new loss function. and in different Predicting the age of organisms.

[0049] different For detailed results, see Figure 4 , Figure 4 (a) For males in different The detailed results below Figure 4 (b) For women in different The detailed results are as follows. Accuracy, Precision, Recall, F1 score, and the area under the individual sample feature curve (AUC) show that smaller... The results are even better. However, the mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), and coefficient of determination are all related. , Indicator too small This may lead to overfitting; among which, It is the arithmetic mean of the ages over time.

[0050] At the same time, when When between 0 and 1, MAE, MSE, RMSE and The changes are relatively small. According to the "elbow rule," in MLR, the area under the difference curve for accuracy, precision, recall, F1 score, and variance suggests... It may reach its optimal state at that time. Figure 4 Loss1 in the text is not introduced. loss function, .

[0051] Evaluation from a statistical perspective for different models: Statistical evaluation metrics primarily measure the fit performance of the prediction model. As shown in Table 1 below, when... When the values ​​are 1, 0, -1, -0.5 and 0.5, the results show that KDM has the best predictive performance, with an MAE value of 4.56 for males and 5.16 for females.

[0052] Table 1. Using different methods Statistical performance evaluation indicators of different models grouped by gender

[0053] Table 2 also assesses the ADC (age difference correlation) bias, defining two additional indicators. )and It should be noted that in many previous studies, the ADC bias has manifested as an overestimation of biological age in younger populations and an underestimation of biological age in older populations.

[0054] Although the predictive performance of these three models is not as good as that of KDM, their results are consistent with previous studies, with MAE values ​​all in the range of 6-8. Furthermore, it can be observed that when the parameters... The model performance stabilized when the value changed from 0 to 1. The MAE for males remained around 7, while the MAE for females remained around 6. However, when... When the value drops below 0, the prediction performance degrades significantly, resulting in lower accuracy. The value is close to or drops below zero.

[0055] Clinical prospect assessment of different models: To further evaluate the predictive model's ability to identify unhealthy individuals in clinical practice, all data were input into the trained model to obtain arterial age. Five evaluation metrics were calculated: precision, accuracy, recall, F1 score, and area under the difference curve, as shown in Table 2. Although the KDM model had relatively low precision (64% for men and 50% for women), its recall was very high (98% for both men and women). This performance characteristic makes it suitable as a preliminary screening tool for unhealthy individuals. From a clinical screening perspective, The smaller the value, the higher the recall rate. For example, The recall rate at -1 is significantly higher than that at 1. The recall rate at 1 indicates that the IRD proposed in this application can effectively enhance the identification of non-high-risk groups. However, as shown in Table 3, when When set to -1 or -0.5, the corresponding The value will become very small, indicating a potential risk of overfitting in the model. It is worth noting that when excluding IRD ( When the value is 1, although a MAE comparable to other models in Table 3 is achieved, the clinical assessment metrics in Table 3 are the worst. In contrast, integrating the IRD into the loss function ( =0) can not only effectively control MAE, but also significantly improve the model's ability to detect unhealthy individuals. For example, MLR in The recall rates at a value of 1 were only 0.49 (female) and 0.56 (male), while... When the value is 0, the recall rates rise to 0.76 and 0.84, respectively. This empirically validates the ability of IRD to enhance the clinical utility of risk prediction.

[0056] Table 2 Clinical Perspective Assessment Based on Different Models of Accuracy, Precision, Recall, and F1 Score

[0057] In summary, this application proposes a novel loss function oriented towards individual differences to correct systematic biases in biological age prediction models. By incorporating individual risk differences (IRDs) into the training objective, it provides a more accurate and clinically relevant assessment of biological age, transcending the limitations of time and age. Results show that the proposed IRD-based loss function not only corrects the systematic overestimation of biological age in younger individuals and the systematic underestimation of biological age in older individuals, but also enhances the model's ability to identify the heterogeneity of biological aging.

[0058] Compared to traditional methods, the IRD loss function significantly improved clinical evaluation metrics such as recall, precision, and AUC. Although performance on conventional statistical metrics like MAE and MSE was similar to traditional methods, the introduction of IRD delivered substantial clinical benefits, particularly in identifying “unhealthy” individuals. These findings highlight the potential and clinical value of integrating individual risk factors into bio-age prediction.

[0059] In summary, although the IRD method performs comparably to traditional methods in terms of traditional statistical indicators, it offers significant improvements in clinical efficacy. Taking a multiple linear regression model as an example, while both the traditional loss function and the IRD loss function maintain a mean absolute error (MAE) of 7–8 for men and 6–7 for women, the IRD loss function increases recall from 0.49 to 0.76 for men and from 0.56 to 0.84 for women; accuracy from 0.50 to 0.63 for men and from 0.51 to 0.62 for women; and AUC from 0.50 to 0.71 for men and from 0.54 to 0.78 for women. This IRD-based loss function not only effectively corrects prediction bias but also validates the rationale behind the IRD decomposition theory, ultimately providing clinicians with a more reliable assessment tool.

[0060] This application deconstructs biological age into two dimensions: chronological age and individual aging deviation, achieving a paradigm shift from "simply fitting chronological age" to "accurately identifying biological anomalies." By constructing a composite loss function, the model, during training, not only focuses on the statistical error between the predicted value and the true age but also emphasizes sensitivity to aging deviations within the same age group. Experimental results show that when the weight coefficients... When the value decreased from 1 to 0, although traditional statistical indicators (such as MAE) did not change significantly, key clinical indicators (such as recall) improved significantly, indicating that the IRD can more effectively capture biological aging characteristics related to health status, especially demonstrating a stronger ability to identify high-risk individuals. This advantage is not only applicable to biomarkers related to arteriosclerosis, but also has the potential to be transferred to multimodal aging assessment tasks such as brain age and heart age. At the same time, the study reveals that a single evaluation perspective may mislead the direction of optimization, for example... The highest recall rate but lowest R² at -1 reflects the trade-off between sensitivity and generalization ability, highlighting the necessity of multi-dimensional performance evaluation. This design provides a theoretical basis for parameter tuning and promotes the evolution of biological age models from "mathematical precision" to "clinical value." Future work requires further refinement of the mathematical expression of the IRD and long-term follow-up validation to enhance the model's robustness and transferability in diverse populations, thus facilitating the clinical translation of biological age assessment.

[0061] Some of the data in the above formula are calculated by removing dimensions and taking their numerical values. The formula is the closest to the real situation obtained by software simulation of a large amount of collected data. The preset parameters and preset thresholds in the formula are set by those skilled in the art according to the actual situation or obtained through simulation of a large amount of data.

Claims

1. A biological age prediction and correction method based on individual differences, characterized in that, include: Multimodal physiological data and corresponding time ages of multiple healthy individual samples are obtained. The physiological data are biological data that objectively reflect the physiological functional state of the human body, and the time age is the actual number of years that the individual sample has experienced since birth. Physiological data are preprocessed to obtain biofeature vectors, which are then input into a biological age prediction model to output the initial biological age prediction value for an individual sample. The biological age prediction value is the age of the actual biological aging degree, which may be higher or lower than the time age. Biological age prediction models are trained based on artificial intelligence neural network models; Based on the time age of individual samples and the corresponding initial biological age prediction value, an average biological age prediction function is obtained by fitting, which represents the average biological age prediction value of the same age group at any time age. Based on the predicted biological age value and the average biological age prediction function, the aging difference value of an individual sample relative to the same age group is calculated. Based on the aging difference value, a composite loss function is constructed. The biological age prediction model is trained with minimizing the composite loss function as the optimization objective to obtain the biological age prediction correction model. The standardized biofeature vector of the individual to be tested is input into the biological age prediction and correction model, and the bias-corrected biological age prediction value is output.

2. The biological age prediction and correction method based on individual differences according to claim 1, characterized in that, The calculation of the aging difference value of an individual relative to their age group includes: Based on the average biological age prediction function, the predicted average biological age value corresponding to the time age is obtained; through the formula... The aging difference value of individual sample i relative to the same age group was calculated. Where i is the sequence number of the individual sample, Let be the predicted biological age of the i-th individual sample. Let be the predicted average biological age of the i-th individual sample.

3. The biological age prediction and correction method based on individual differences according to claim 2, characterized in that, The construction of the composite loss function based on the aging difference value includes: By incorporating the aging difference value into the traditional MSE loss function, a composite loss function is obtained. : ； in, The optimal hyperparameters are used to adjust the weights of the IRD in the loss function, where n is the total number of individual samples. Let be the age of the i-th individual sample over time. Let be the predicted biological age of the i-th individual sample.

4. The biological age prediction and correction method based on individual differences according to claim 3, characterized in that, The optimal hyperparameters The methods for determining this include: The hyperparameter search range is preset, and multiple candidate values ​​are selected within the preset search range; For each candidate value, a corresponding biological age prediction model is trained using the training data, and the statistical performance indicators and clinical performance indicators of the model on the validation set are calculated respectively. The statistical performance indicators include mean absolute error (MAE), mean squared error (MSE), root mean squared error (RMSE), and coefficient of determination. At least one of the following clinical performance metrics, wherein the clinical performance metrics include at least one of recall, accuracy, F1 score, and area under the characteristic curve (AUC) for individual samples; Based on the preset clinical application goals, different weights are assigned to each performance indicator, and the weighted sum of the performance indicators corresponding to each candidate value is calculated to obtain a comprehensive score; among them, the candidate value corresponding to the highest comprehensive score is the optimal hyperparameter.

5. The biological age prediction and correction method based on individual differences according to claim 4, characterized in that, The construction of the characteristic curves of the individual samples includes: Analyze the health risk score of each individual sample, and use the health risk score of each individual sample as a classification threshold in turn. For each classification threshold, predict the label category for each individual sample; where individual samples with a classification threshold greater than the preset classification threshold are labeled as high-risk, and the remaining individual samples are labeled as low-risk. Based on the predicted label category and the actual label category for each individual, the number of true positive cases (TP), false positive cases (FP), true negative cases (TN), and false negative cases (FN) are calculated. Calculate the true positive rate (TPR) = TP / (TP + FN) and the false positive rate (FPR) = FP / (FP + TN) for each individual sample; construct the characteristic curve of the individual sample with FPR as the x-axis and TPR as the y-axis.

6. The biological age prediction and correction method based on individual differences according to claim 5, characterized in that, The true positive count (TP) is the number of individual samples whose actual and predicted label categories are both high risk; the false positive count (FP) is the number of individual samples whose actual label category is low risk and whose predicted label category is high risk; the true negative count (TN) is the number of individual samples whose actual and predicted label categories are both low risk; and the false negative count (FN) is the number of individual samples whose actual label category is high risk and whose predicted label category is low risk.

7. The biological age prediction and correction method based on individual differences according to claim 5, characterized in that, The analysis of the health risk score for each individual sample includes: Through formula The health risk score for each individual sample was calculated. ;in, For unit difference fractions, when ,but The value is a positive integer greater than 0. ,but The value is a positive integer less than 0.

8. The biological age prediction and correction method based on individual differences according to claim 1, characterized in that, The method for obtaining the average biological age prediction function includes: Individual samples are divided into several intervals based on their time age. Within each interval, the arithmetic mean of the initial biological age predictions of all samples is calculated to obtain the average biological age prediction value for the corresponding interval. The median of the time age in each interval and the corresponding average biological age prediction value constitute several discrete points. Arbitrary time ages are inserted between each discrete point, and the corresponding average biological age prediction value is calculated through linear interpolation. The discrete points are fitted to obtain a continuous average biological age prediction function. The width of each interval is less than a preset width.

9. The biological age prediction and correction method based on individual differences according to claim 1, characterized in that, The preprocessing of physiological data to obtain biological feature vectors includes: Missing values ​​were processed in the multimodal physiological data collected from individual sample i; Biometrics with absolute Pearson correlation coefficients greater than a preset correlation threshold were selected as valid features. For each valid biometric, its mean and standard deviation were calculated, and then analyzed using the formula... The standardized value of the j-th biometric item for the i-th individual sample was calculated. j represents the sequence number of the biometric indicator; The standardized biological indicators are constructed into biological feature vectors in a preset order to obtain the biological feature vector of the i-th individual.

10. A biological age prediction and correction system based on individual differences, operating based on the biological age prediction and correction method based on individual differences according to any one of claims 1-9, characterized in that, It includes a prediction module, as well as an acquisition module and a correction module connected to it; The acquisition module is used to acquire multimodal physiological data and corresponding time ages of multiple healthy individual samples; The prediction module is used to preprocess physiological data to obtain biofeature vectors, which are then input into the biological age prediction model to output the initial biological age prediction value of individual samples. Based on the time age of individual samples and the corresponding initial biological age prediction value, the average biological age prediction function is fitted to obtain the mean biological age prediction function. The correction module is used to calculate the aging difference value of an individual sample relative to the same age group based on the predicted biological age value and the average biological age prediction function. Based on the aging difference value, a composite loss function is constructed. The biological age prediction model is trained with minimizing the composite loss function as the optimization objective to obtain a bias-corrected biological age prediction model. The standardized biofeature vector of the individual to be tested is input into the biological age prediction and correction model, and the bias-corrected biological age prediction value is output.

Images