Training set feature screening method, immune age prediction method and device

Abstract

The invention provides a training set feature screening method for immune age prediction, and an immune age prediction method and device. The training set feature screening method comprises the following steps: obtaining immune index data sets of a plurality of individuals; performing feature expansion on the immune index data sets of the plurality of individuals by adopting a depth feature synthesis algorithm; on the basis of the expanded feature set, a Boruta algorithm is adopted to verify the importance of the features on the immune age, and a stable related feature set is obtained through screening; through a forward selection algorithm, the immune age prediction performance of the model is taken as a constraint condition, the stable related feature set is optimized and screened, and a key feature set with the minimum number is obtained on the premise that the predetermined prediction precision is ensured and is taken as immune age key features to construct a training set.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics processing technology, and in particular to a method for screening training set features, an immune age prediction method, and an apparatus for immune age prediction. Background Technology

[0002] The functional decline of the immune system with age, known as immunosenescence, is closely related to a decline in an individual's ability to respond to infections and vaccines, as well as the development of age-related diseases such as cancer and autoimmune diseases. Therefore, accurately quantifying and assessing an individual's "immune age" is of great value for early warning of health risks and guiding personalized health management.

[0003] Currently, the assessment of immune age mainly relies on the detection and analysis of a series of immunological indicators. Traditional methods are usually based on expert knowledge or simple statistical analysis, selecting several known age-related immune indicators, such as the proportion of specific lymphocyte subsets, and performing linear combinations or constructing simple scoring models. However, these methods have significant limitations: First, the immune system is a highly complex network, and its aging process involves multi-level, non-linear changes from cell number and subset proportions to cell functional states (such as metabolism, activation, and exhaustion), and relying on only a few pre-set indicators is insufficient to fully capture its complexity; Second, massive amounts of high-dimensional immunoassay data contain a wealth of potential information, such as multicolor flow cytometry which can simultaneously detect dozens of indicators, and traditional methods lack effective automated means to extract key feature combinations that can robustly characterize the aging process; Third, complex interaction effects may exist between different immune indicators, and traditional linear models are unable to effectively model such non-linear relationships, resulting in limited prediction accuracy and biological interpretability.

[0004] In recent years, machine learning methods have been applied to bioinformatics. Some studies have attempted to model immune indicators using conventional statistical methods or basic machine learning algorithms. However, these methods still face challenges when applied to immune age prediction: on the one hand, they often directly use the original detection indicators without fully considering how to automatically construct and screen the most predictive and robust high-order feature combinations from high-dimensional data that may exhibit multicollinearity; on the other hand, in the feature selection stage, there is a lack of a systematic selection process that can balance the relevance of features to the target, the importance and stability of the features themselves, and the simplicity of the final model. This can easily lead to model overfitting or redundant and unstable selected feature sets, thereby affecting the model's generalization ability and practical application value.

[0005] Therefore, there is an urgent need to provide a method for predicting immune age. Summary of the Invention

[0006] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a training set feature screening method, immune age prediction method and device for immune age prediction, which can extract a set of key features characterizing immune aging from the original immunological indicators, and make immune age prediction more accurate and faster based on this.

[0007] To achieve the above objectives, the main technical solutions adopted by the present invention include:

[0008] In a first aspect, the present invention provides a training set feature selection method for immune age prediction, comprising the following steps: obtaining immune indicator datasets of multiple individuals; expanding the immune indicator datasets of multiple individuals using a deep feature synthesis algorithm; verifying the importance of features with respect to immune age using the Boruta algorithm based on the expanded feature set, and selecting a stable and relevant feature set; optimizing and selecting the stable and relevant feature set using a forward selection algorithm with the immune age prediction performance of the model as a constraint, and obtaining the set of key features with the smallest number of features while ensuring a predetermined prediction accuracy as key features for immune age to construct the training set.

[0009] Optionally, the immune indicator dataset includes immune cell subsets and mitochondrial function indicators obtained by flow cytometry, as well as routine hematological indicators obtained by blood analysis.

[0010] Optionally, a deep feature synthesis algorithm is used to extend the features of the immune indicator datasets of multiple individuals. This includes: performing addition, subtraction, multiplication, and division on the immunological features in the immune indicator dataset belonging to a single individual, and generating higher-order combined features in a recursive combination manner. Specifically, when performing multiplication and division arithmetic primitive operations, the immunological features are first logarithmized, and then addition is used to replace multiplication, and subtraction is used to replace division, so as to generate equivalent combined features.

[0011] Optionally, based on the expanded feature set, the Boruta algorithm is used to verify the importance of features with respect to immune age, and a stable relevant feature set is obtained by screening. This includes: firstly, using the mutual information algorithm to screen the expanded feature set to obtain a preliminary feature set that has a statistical dependence on age, and then using the Boruta algorithm to verify the importance of the preliminary feature set with respect to immune age, and then screening to obtain a stable relevant feature set.

[0012] Optionally, a forward selection algorithm is used to optimize and filter the stable relevant feature set under the constraint of the model's immune age prediction performance to obtain the key feature set with the smallest number of features while ensuring the predetermined prediction accuracy. This includes: gradually constructing a feature subset from an empty set through forward iteration; in each iteration, selecting a feature that can maximize the improvement of the immune age prediction model performance from the stable relevant features that have not yet been selected into the feature subset; continuing the iteration until the model performance gain brought by the newly added feature is lower than the preset performance gain threshold; and determining the feature subset that has been constructed before the last iteration as the key feature set for immune age.

[0013] Secondly, the present invention provides an immune age prediction method, comprising the following steps: acquiring an immune indicator dataset of the individual to be tested; performing feature expansion processing on the immune indicator dataset using a deep feature synthesis algorithm to obtain an expanded feature set; selecting a corresponding feature subset from the expanded feature set as model input features based on pre-determined key immune age features; inputting the input features into a pre-trained immune age prediction model; and outputting the predicted immune age value for the corresponding individual; wherein, the key immune age features are determined using the training set feature selection method described above; and the immune age prediction model is trained based on a training set constructed from the key immune age features.

[0014] Optionally, the pre-trained immune age prediction model is an ensemble model optimized using a Bagging ensemble learning strategy;

[0015] The training process of the immune age prediction model includes: performing multiple bootstrap samplings with replacement on the training dataset to obtain multiple training subsets; and training a base prediction model independently using each training subset to obtain multiple base prediction models.

[0016] The input features are fed into the immune age prediction model, which involves feeding them into multiple base prediction models. The immune age prediction values ​​output by the multiple base prediction models are then averaged to obtain the final immune age prediction value.

[0017] Optionally, the immune age prediction model is a machine learning model, wherein the machine learning model is any one of the following: random forest model, lightweight gradient boosting tree model, extreme gradient boosting tree model, K-nearest neighbor model, and support vector machine model.

[0018] Optionally, the immune age prediction model is a tree-based machine learning model, wherein the tree-based machine learning model is any one of the random forest model, the lightweight gradient boosting tree model, and the extreme gradient boosting tree model.

[0019] Thirdly, the present invention provides an immune age prediction device, comprising:

[0020] The acquisition module is used to acquire the immune indicator dataset of the individual to be tested;

[0021] The feature extraction module is used to perform feature expansion processing on the immune index dataset using a deep feature synthesis algorithm to obtain an expanded feature set. Based on the pre-determined key features of immune age, the module selects a corresponding feature subset from the expanded feature set as the input features of the model.

[0022] The prediction module is used to input the input features into a pre-trained immune age prediction model and output the predicted immune age value for the corresponding individual.

[0023] Among them, the key features of immune age are determined by the training set feature selection method described above; the immune age prediction model is trained based on the training set constructed from the key features of immune age.

[0024] The beneficial effects of this invention are:

[0025] This invention provides a training set feature selection method and an immune age prediction method for immune age prediction. These methods establish a comprehensive data foundation based on multi-level immune system indicators and innovatively employ a deep feature synthesis algorithm to automatically mine complex nonlinear relationships and interaction effects between indicators, overcoming the limitations of traditional methods that rely on limited preset indicators and linear models. Furthermore, the core feature selection process, which pre-validates using the Boruta algorithm and optimizes forward selection, automatically and with high confidence selects a concise set of key features that contribute stably to age prediction from a massive dataset. This significantly reduces model complexity and the risk of overfitting while enhancing the model's interpretability and biological rationality. Finally, based on the selected key features, a pre-trained prediction model is used to quickly and accurately output an individual's immune age, providing a reliable technical solution for efficient and robust quantitative assessment of immune status. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating the immune age prediction method according to a specific embodiment of the present invention.

[0027] Figure 2 A graph showing the relationship between 61 immune cell subsets and mitochondrial function indicators according to a specific embodiment of the present invention.

[0028] Figure 3a For the selection and accumulation of original features R 2 Changes in a dual-axis graph;

[0029] Figure 3b To stabilize the screening and accumulation of features in the relevant feature set R 2 Changes in a dual-axis graph;

[0030] Figure 4 The bee colony diagram represents the SHAP values ​​that characterize the impact of features on the model output. In the diagram, MMP represents the MMP value. low The abbreviation of %. Detailed Implementation

[0031] To better explain and facilitate understanding of the present invention, it will be described in detail below with reference to the accompanying drawings and specific embodiments. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a clearer and more thorough understanding of the invention and to fully convey the scope of the invention to those skilled in the art.

[0032] The age distribution of the samples showed that the healthy control group (HC) samples were mainly concentrated in the adolescent stage, while the tumor patient (TM) samples were mainly distributed in the middle-aged and elderly stages, indicating a significant difference in age structure. In the correlation analysis between immune indicators and age, we calculated the Pearson correlation coefficient between each feature and age, and set a threshold of 0.3 as the standard for significant correlation. The results showed that the proportion of cell subsets such as T4Tem and T8Tem was significantly positively correlated with age, while the proportion of cells such as T4Tn, T8Tn, and B cells decreased significantly with increasing age. Regarding mitochondria, mitochondrial mass of most cell types was negatively correlated with age, and the proportion of mitochondrial low membrane potential was generally negatively correlated with age, indicating that cell activity and functional status may weaken with increasing age. The expression ratio of PD-1 protein did not show a significant correlation. Since the Pearson correlation coefficients of most features were below 0.3, it indicates that the linear relationship between them and age is not significant. This suggests that immune features may reflect age-related changes in a non-linear or interactive manner. Based on this, the present invention employs a feature combination and interaction construction method to uncover more complex immunological feature relationships, thereby improving the model's ability to represent immune age.

[0033] like Figure 1 As shown, the present invention provides a method for predicting immune age, comprising the following steps:

[0034] 101. Obtain the immune indicator dataset of the individual to be tested.

[0035] Specifically, the immune indicator dataset consists of human immune indicator data detected from individual EDTA-anticoagulated peripheral blood samples using the TBNK mitochondrial detection kit (UBBiotech, Cat#UB32229) and the TFUN mitochondrial detection kit (UBBiotech, Cat#UB32479, Hunan Medical Device Registration Certificate 20222401162) based on the MitoDye mitochondrial probe, on a flow cytometry platform with a configuration of dual lasers and six or more colors.

[0036] As an example, the immune marker dataset includes 61 immune cell subsets and mitochondrial function indicators obtained by flow cytometry, and 11 routine hematological indicators obtained by blood analysis.

[0037] Among them, such as Figure 2 As shown, 61 immune cell subsets and mitochondrial function indicators include: white blood cell count (WBC%), white blood cell mitochondrial mass (total fluorescence intensity) (WBC.MM(T), and lymphocyte mitochondrial low membrane potential (Lymph.MMP). low %, Lymphocyte mitochondrial mass (total fluorescence intensity) Lymph.MM(T), percentage of lymphocytes Lymph%, Monocyte mitochondrial mass (total fluorescence intensity) Mono.MM(T), percentage of monocytes Mono%, B cell mitochondrial low membrane potential percentage B.MMP low %, B cell mitochondrial mass (total fluorescence intensity) B.MM(T), B cell percentage B%, NK cell mitochondrial low membrane potential percentage NK.MMP low %, NK cell mitochondrial mass (total fluorescence intensity) NK.MM(T), NK cell percentage NK%, T cell mitochondrial low membrane potential percentage T.MMP low %, T cell mitochondrial mass (total fluorescence intensity) T.MM(T), T cell percentage T%, total T cells (CD3+) + ) Mitochondrial low membrane potential ratio CD3.MMP low %, Total T cell functional subset mitochondrial mass (for functional subsets with normal mitochondrial membrane potential) T.MM(F), Total T cells (CD3 + ) percentage CD3 + %, the proportion of helper T cells with low mitochondrial membrane potential (Th.MMP) low %, helper T cell mitochondrial mass (total fluorescence intensity) Th.MM(T), percentage of helper T cells Th%, helper T cells (CD4+) + ) Mitochondrial low membrane potential ratio CD4 MMP low %, mitochondrial mass of helper T cell functional subsets Th.MM(F), helper T cells (CD4) + ) percentage CD4+ %, the proportion of mitochondrial low membrane potential in cytotoxic T cells (Ts.MMP) low %, cytotoxic T cell mitochondrial mass (total fluorescence intensity) Ts.MM(T), percentage of cytotoxic T cells Ts%, cytotoxic T cells (CD8) + ) Mitochondrial low membrane potential ratio CD8.MMP low %, mitochondrial mass Ts.MM(F) of cytotoxic T cell functional subsets, cytotoxic T cells (CD8) + ) percentage CD8 + %, initial CD4 + The proportion of T cell mitochondrial low membrane potential (T4Tn.MMP) low %, initial CD4 + T cell functional subset mitochondrial mass T4Tn.MM(F), initial CD4 + T cell percentage (T4Tn%), initial CD8 + The proportion of T cell mitochondrial low membrane potential (T8Tn.MMP) low %, initial CD8 + T cell functional subset mitochondrial mass T8Tn.MM(F), initial CD8 + T cell percentage (T8Tn%), effector CD4 + The proportion of T cell mitochondrial low membrane potential (T4Tef.MMP) low %, Effect memory CD4 + T cell PD-1 expression ratio (T4Tef.PD-1) + %, effect CD4 + T cell functional subsets mitochondrial mass T4Tef.MM(F), effector CD4 + T cell percentage (T4Tef%), effector CD8 + The proportion of T cell mitochondrial low membrane potential (T8Tef.MMP) low %, effect CD8 + T cell PD-1 expression ratio T8Tef.PD-1 + %, effect CD8 + T cell functional subsets mitochondrial mass T8Tef.MM(F), effector CD8 + T cell percentage (T8Tef%), central memory CD4 + The proportion of T cell mitochondrial low membrane potential (T4Tcm.MMP) low %, Central Memory CD4 + T cell PD-1 expression ratio (T4Tcm.PD-1) + %, Central Memory CD4 + T cell functional subsets mitochondrial mass T4Tcm.MM(F), central memory CD4+ T cell percentage (T4Tcm%), effector memory (CD4) + The proportion of T cell mitochondrial low membrane potential (T4Tem.MMP) low %, Effect memory CD4 + T cell PD-1 expression ratio (T4Tem.PD-1) + %, effect memory CD4 cell functional subset mitochondrial mass T4Tem.MM(F), effect memory CD4 + T cell percentage (T4Tem%), central memory CD4 + The proportion of T cell mitochondrial low membrane potential (T8Tcm.MMP) low %, Central Memory CD4 + T cell PD-1 expression ratio (T8Tcm.PD-1) + %, Central Memory CD4 + T cell functional subsets mitochondrial mass T8Tcm.MM(F), central memory CD4 + T cell percentage (T8Tcm%), effector memory (CD4) + The proportion of T cell mitochondrial low membrane potential (T8Tem.MMP) low %, Effect memory CD4 + T cell PD-1 expression ratio (T8Tem.PD-1) + %, mitochondrial mass T8Tem.MM(F) of the CD4 functional subset of cells with effect memory, and CD4 effect memory + T cell percentage (T8Tem).

[0038] The 11 routine hematological indicators include white blood cell count (WBC), red blood cell count (RBC), hemoglobin concentration (HGB), hematocrit (HCT), mean corpuscular volume (MCV), red blood cell distribution width (RDW), mean corpuscular hemoglobin (MCH), mean corpuscular hemoglobin concentration (MCHC), platelet count (PLT), plateletcrit (PCT), and mean platelet volume (MPV). These indicators reflect the body's hematopoietic function, oxygen transport capacity, coagulation, and immune status.

[0039] The overall structure of the above 72 immune indicators reflects the hierarchical distribution of the immune system from total white blood cells to cell subsets and their functional parameters, providing a systematic indicator framework for modeling immune aging and immune age.

[0040] 102. The immune index dataset is expanded using a deep feature synthesis algorithm to obtain an expanded feature set; based on the pre-determined key features of immune age, a corresponding feature subset is selected from the expanded feature set as the input features of the model.

[0041] Preferably, a deep feature synthesis algorithm is used to extend the features of the immune indicator dataset, including performing four arithmetic primitive operations (addition, subtraction, multiplication, and division) on the immunological features to generate higher-order combined features through recursive combination. Specifically, when performing multiplication and division primitive operations, the immunological features are first logarithmically processed, and then addition is used to replace multiplication, and subtraction is used to replace division, to generate equivalent combined features. By recursively combining basic features, a more complex feature hierarchy structure is formed, thereby capturing the nonlinear and interactive relationships between the original variables. Furthermore, when performing multiplication and division operations, they are converted into addition and subtraction operations after logarithmic processing, rather than directly multiplying or dividing the original data, which prevents the data from being affected by maxima and minima.

[0042] As an example, a deep feature synthesis algorithm is used to expand the features of 72 immune indicators (performing one deep feature synthesis), resulting in an expanded feature set containing 4458 features. This greatly expands the model's input space.

[0043] 103. Input the input features into the pre-trained immune age prediction model and output the predicted immune age value of the individual to be tested.

[0044] The immune age prediction model is trained on a training set constructed from key features of immune age.

[0045] Preferably, the pre-trained immune age prediction model is an ensemble model optimized using a Bagging ensemble learning strategy. The training process of the immune age prediction model includes: performing multiple bootstrapping operations with replacement on the training dataset to obtain multiple training subsets; independently training a base prediction model using each training subset to obtain multiple base prediction models; inputting the input features into the immune age prediction model (specifically, inputting them into the multiple base prediction models); and averaging the immune age prediction values ​​output by the multiple base prediction models to obtain the final immune age prediction value. In this way, the immune age prediction model exhibits strong generalization ability and stability, and achieves high accuracy in immune age prediction.

[0046] It should be noted that the training dataset used in the training of the immune age prediction model can be the same training dataset used in the determination of key features of immune age described below.

[0047] The immune age prediction model is a machine learning model, specifically, any one of the following: Random Forest, LightGBM, XGBoost, KNN, and SVM. More specifically, any one of these models serves as the base prediction model.

[0048] By comparing the performance of the random forest model, lightweight gradient boosting tree model, extreme gradient boosting tree model, K-nearest neighbor model, and support vector machine model as immune age prediction models, the tree-based machine learning model shows better performance, with the random forest model exhibiting the best performance. Therefore, preferably, the immune age prediction model is a tree-based machine learning model, wherein the tree-based machine learning model is any one of the random forest model, lightweight gradient boosting tree model, and extreme gradient boosting tree model. More preferably, the immune age prediction model is a random forest model.

[0049] The process of determining the key characteristics of immune age is explained in detail below.

[0050] The process of determining key characteristics of immune age includes the following steps:

[0051] 1001. Obtain immune indicator data from multiple individuals.

[0052] The immune index data for each individual includes 72 immune indicators, namely 61 immune cell subsets and mitochondrial function indicators obtained by flow cytometry, and 11 routine hematological indicators obtained by blood analysis.

[0053] Preferably, the method further includes: preprocessing immune indicator data from multiple individuals. The preprocessing process includes: using the interquartile range method to detect outliers, removing non-numerical values, infinity, and unreasonable negative values, while retaining statistically significant outliers. In immunological research, extreme values ​​often reflect physiological differences or potential pathological signals between individuals, rather than measurement errors. Therefore, not removing statistically significant outliers (i.e., extreme values) ensures the validity and consistency of the data.

[0054] 1002. A deep feature synthesis algorithm is used to expand the features of immune index data of multiple individuals.

[0055] The process of feature expansion is the same as step 102 above, and will not be repeated here.

[0056] It should be noted that, to further expand the features, we manually constructed features based on prior knowledge in the field of immunometabolology to further explore potential deep biological associations. For example, based on the known biological mechanisms that "mitochondrial membrane potential reflects cellular metabolic activity" and "PD-1 signaling pathway participates in T cell metabolic reprogramming," we can artificially design interactive features that combine cell subset proportions, mitochondrial function indicators, and PD-1 expression levels, aiming to explore their potential compound effects in age-related changes in immune function. However, model evaluation results show that high-order manual features based on domain knowledge did not improve the overall predictive performance.

[0057] 1003. Based on the expanded feature set, the Boruta algorithm is used to verify the importance of features with respect to immune age, and a stable and relevant feature set is obtained by screening.

[0058] Specifically, the Boruta algorithm for validation and selection includes the following steps: For each original feature in the expanded feature set, a corresponding random control feature is generated by randomizing its numerical order. All original features and their corresponding random control features are merged into a mixed feature set. Using this feature set as input, multiple rounds of iterative computation are performed based on a random forest model. In each iteration, the importance score of each feature in the feature set for predicting immune age is calculated and recorded. After multiple rounds of iteration, for each original feature, its importance score obtained in the multiple iterations is statistically analyzed and compared with the importance scores obtained by its corresponding random control feature in the multiple iterations. Original features with consistently high importance scores and statistically significantly higher than their corresponding random control features are selected. The set of these validated original features is determined as the stable relevant feature set. By introducing random control features as a "noise baseline," it is ensured that the contribution of the finally selected features to age prediction is stable and realistic, rather than a random result.

[0059] Preferably, based on the expanded feature set, the Boruta algorithm is used to verify the importance of features with respect to immune age, and a stable relevant feature set is obtained through screening. This includes: firstly, using the mutual information algorithm to determine the mutual information between each feature in the expanded feature set and age. This indicator can effectively measure the linear and non-linear statistical dependencies between features and age. Features with mutual information higher than a preset threshold are selected to form an initial feature set with a strong statistical dependency on age. Subsequently, the Boruta algorithm screening and verification steps described above are performed on the initial feature set to obtain a stable relevant feature set. This two-step strategy first uses the mutual information algorithm to quickly eliminate a large number of obviously irrelevant features, narrowing the candidate range; then, the Boruta algorithm is used to rigorously verify the statistical importance of the simplified candidates, effectively reducing computational complexity while ensuring the accuracy and robustness of the screening.

[0060] 1004. Using the forward selection algorithm, with the model's immune age prediction performance as a constraint, the stable relevant feature set is optimized and screened to obtain the key feature set with the smallest number of features while ensuring the predetermined prediction accuracy, which is then used as the key features for immune age.

[0061] Specifically, a feature subset is gradually constructed starting from an empty set through forward iteration. In each iteration, a feature that can maximize the performance of the immune age prediction model is selected from the stable related features that have not yet been selected into the feature subset. The iteration continues until the model performance gain brought by the newly added feature is lower than the preset performance gain threshold. The feature subset that has been constructed before the last iteration is determined as the key feature set for immune age.

[0062] After determining the key characteristics of immune age, this invention conducted an in-depth analysis of the feature screening results to verify the effectiveness of the screening process and elucidate its scientific implications. The specific process is as follows:

[0063] Using a forward selection algorithm, with the model's immune age prediction performance as a constraint, the original features (i.e., 72 immune features) were optimized and filtered to obtain the set of key features with the smallest number while ensuring the model's predetermined prediction accuracy. This set was then used as a comparative reference.

[0064] Figure 3a For the selection and accumulation of original features R 2 Changes in a dual-axis graph; Figure 3b To stabilize the screening and accumulation of features in the relevant feature set R 2 A dual-axis graph showing changes. Figure 3a and Figure 3b In the image, a bar chart shows the ranking of feature importance. Red bars represent features ultimately selected by the forward selection algorithm, blue bars represent features that were not selected by the forward selection algorithm, and the green curve and its shaded area represent the model's coefficient of determination R when using the forward successive selection algorithm. 2 The cumulative changes as features are introduced one by one and their fluctuation range in cross-validation.

[0065] like Figure 3a As shown, when filtering the original features, the R-value of the model decreases as important features are gradually introduced. 2 The value rises rapidly. After incorporating approximately 13 original features ranked by importance, the model's R-value increases rapidly. 2 The value stabilized, reaching approximately 0.766. Afterward, the performance gain (R²) of adding new features continued to increase. 2The improvement becomes negligible. This phenomenon clearly indicates that these 13 features already contain the core information needed to predict immune age, and subsequent features may contain redundant information or noise.

[0066] like Figure 3b As shown, when filtering features from a stable and relevant feature set, the R-value of the model decreases as important features are gradually introduced. 2 The value rises rapidly. After including approximately 11 features ranked by importance, the model's R-value increases rapidly. 2 The value stabilized at approximately 0.773, slightly outperforming the model using 13 original features. This indicates that the combined features constructed through the feature engineering of this invention can carry the same or even stronger predictive information with fewer features. Based on this, this invention identifies these 11 combined features as the final core feature set for constructing a high-precision immune age prediction model.

[0067] Comparative analysis of the 13 original features and 11 stable related features revealed a significant overlap in their composition. For example, T8Tn% and T8Tcm.MMP low The percentage and other indicators play a significant role in both groups of features. This overlap indicates that the feature screening process of this invention can reliably identify key signals that are biologically closely related to immune aging and have repeatability, verifying the robustness and biological rationality of the screening process.

[0068] A key discovery highlights the value of the feature engineering (i.e., deep feature synthesis) of this invention. For example... Figure 3a and Figure 3bAs shown, the two original features, T4Tcm.MM(F) and T4Tn.MM(F), rank relatively low in importance individually. However, when they are combined into a single feature through subtraction, such as T4Tcm.MM(F) - T4Tn.MM(F), the importance of this new feature jumps to the top five among all stable and relevant features. This phenomenon indicates that in complex immune system data, the absolute level of a single feature may have limited predictive power, while the relative relationships, differences, or proportions (i.e., interaction information) between different features often contain stronger predictive signals. Deep feature synthesis algorithms automatically capture such high-order interaction relationships between features through mathematical operations, thereby "amplifying" potential biological signals and demonstrating their advantages in information fusion. The aforementioned leap in the importance of the combined feature may reveal an "imbalance" or "widening gap" in energy metabolism levels among T cell subsets with different functional states during immune aging. This relative change is more sensitive to the functional decline of the immune system than the absolute metabolic level of a single cell subset, thus exhibiting stronger explanatory power in the model. This demonstrates that the feature construction and screening strategy employed in this invention can effectively capture complex biological signals that closely reflect the systemic nature of immune aging.

[0069] The above analysis, through the accompanying figures and data, fully verifies the scientific validity and effectiveness of the "deep feature synthesis extension + Boruta and forward selection two-step screening" technical path of this invention. It not only yields a high-performance, low-redundancy prediction model, but also deepens the understanding of the biological characterization of immune aging.

[0070] After identifying key features of immune age and completing model construction, this invention introduces the SHAP (Shapley Additive Explanations) method to perform interpretability analysis on the final model in order to gain a deeper understanding of the model's decision-making logic and verify the biological significance of the selected features. SHAP analysis can quantify the marginal contribution of each feature to a specific prediction result, thereby transforming the "black box" model into an interpretable decision-making process.

[0071] Figure 4 A beehive diagram representing the SHAP values ​​of features on the model output. For example... Figure 4 As shown, the SHAP beehive diagram visually illustrates the impact of each key feature on the final immune age prediction model output. Each point in the diagram represents a sample, and its position on the X-axis (SHAP value) indicates the specific contribution of that feature to the prediction: being to the right of zero (positive SHAP value) indicates that the feature contributes to a higher predicted immune age, while being to the left (negative SHAP value) indicates a lower predicted immune age. The color of the point represents the original value of that sample on that feature, with red indicating a high value and blue indicating a low value.

[0072] analyze Figure 4 This clearly reveals the stable association between different features and the direction of immunosenescence. For example, for the combined feature ln(T8Tn%) + ln(T8Tcm.MMP) low The figure shows that its high eigenvalues ​​(red dots) are concentrated to the left of the zero point of the SHAP value. This indicates that when the value of this combined feature is high, the model tends to predict a lower immune age, meaning that this key feature is strongly correlated with the "delayed immune aging" signal. Similarly, CD3.MMP low %-T8Tn.MMP low The combined features such as %, T4Tcm.MM(F) – T4Tn.MM(F) also showed the same trend. Other key features exhibited clear "pro-aging" signals. For example, the high eigenvalues ​​(red dots) of the combined feature ln(T4Tem%) + ln(T8Tcm%) were mainly concentrated to the right of the zero point of the SHAP value. This indicates that when this eigenvalue increases, the model predicts a higher immune age, i.e., it is associated with the "accelerated immune aging" process. Combined features MCV + MCHC, CD4 + %+T4Tcm% and others also showed similar pro-aging effects.

[0073] The SHAP analysis results described above, from a data-driven perspective, strongly validate that the key features screened using the method of this invention have clear biological significance. The identification of features that delay aging (such as those involving the combination of naive T cells and functionally normal mitochondria) and those that accelerate aging (such as those involving the combination of effector memory cells and abnormal erythrocyte parameters) is entirely consistent with fundamental understanding in the field of immunosenescence. This demonstrates that the model of this invention is not only a mathematical fitting tool, but also a computational framework capable of capturing and quantifying known biological processes of immunosenescence.

[0074] SHAP analysis not only provides a global perspective but also allows for the interpretation of predictions for any individual test subject. By calculating the SHAP value for each individual across all key features, a "contribution waterfall plot" can be generated, clearly showing whether each feature pushes up or down the individual's predicted age, and by how much. This makes every prediction of the model transparent and traceable, greatly enhancing its credibility and practicality in scientific research and clinical decision support. Doctors or researchers can use this to understand why a particular individual was predicted to have "advanced immune age," and which specific immune indicators contributed the most to the abnormality, thus providing direct clues for subsequent precise interventions.

[0075] Combination Figure 4Through SHAP analysis, this invention not only achieves high-precision prediction of immune age, but also makes the decision-making basis of the model clearly visible through interpretable artificial intelligence technology. This dual advantage ensures that the immune age assessment results provided by this invention are both reliable (high precision) and credible (interpretable and biologically consistent), laying a solid technical foundation for its application in aging mechanism research, health status assessment, and risk warning.

[0076] After completing model construction and interpretability analysis, a comprehensive evaluation of the model's predictive performance, generalization ability, and optimization effect is necessary. The evaluation strategy of this invention strictly distinguishes between internal and external validation to objectively reflect the model's practical application potential.

[0077] Because there are significant differences in age distribution and immune status between healthy individuals (HC) and cancer patients (TM), this invention employs the following strategy to ensure the model learns robust physiological aging patterns: Healthy individuals' data are used as the primary training and testing set, allowing the model to fully learn the changing patterns of immune characteristics from adolescents to young adults. Cancer patient data is used as an independent external validation set to test the model's predictive robustness and biological rationality when facing heterogeneous populations with significantly abnormal immune status. We anticipate and accept that the model's predictive accuracy will decrease on the TM dataset, but we focus on whether it can systematically capture the prevalent trend of "accelerated immune aging" (i.e., predicted age > actual age) in cancer patients, using this as important biological evidence of the model's effectiveness.

[0078] To establish a performance baseline and validate the value of feature engineering, we first input the original 72 immune indicators into five different regression models for training and evaluation. As shown in Table 1, the baseline performance of each model differed on the healthy population test dataset (HC Testdataset). Subsequently, we compared the performance of the model constructed using the stable relevant feature set (20 Features) selected using the method of this invention with the model constructed using the 11 key features finally determined.

[0079] The results show that using only 11 key features, most models (especially tree models) show improved performance on the HC test set. Taking random forest as an example, using 11 key features, its MAE is 5.830 and R² is 0.789, both of which are better than the performance of using all original features (MAE 6.916 and R² 0.758). This proves that the feature selection process of this invention can significantly reduce the number of required detection indicators (by about 85%) while extracting more predictive information, effectively improving the model's prediction efficiency and accuracy.

[0080] A comprehensive comparison of five models reveals that tree-based algorithms (Random Forest, LightGBM, and XGBoost) significantly outperform models such as K-Nearest Neighbors (KNN) and Support Vector Machines (SVM) across all metrics. For example, with 11 features, the R² of the tree model is generally above 0.74, while the R² of KNN and SVM is below 0.3. Therefore, this invention preferentially selects the tree model as the basis for subsequent in-depth optimization.

[0081] Table 1

[0082]

[0083] After selecting high-performing tree models (Random Forest, LightGBM, XGBoost), we further optimized them using a Bagging ensemble learning strategy to improve their stability and generalization ability. The optimization process avoided complex hyperparameter tuning (such as grid search) that could easily lead to overfitting on small sample data; instead, we used the default parameters of each model to ensure its robustness.

[0084] The optimization results are shown in Table 2, which demonstrates the final performance of the three tree models on the HC test set and the TM external validation set after optimization using the Bagging strategy:

[0085] Table 2

[0086]

[0087] After optimization, the model's prediction accuracy was further consolidated and improved. For example, the MAE of the Random Forest model decreased from 5.830 to 5.821, and the R² increased from 0.789 to 0.799; the optimization effect of LightGBM was even more significant. This demonstrates that the Bagging ensemble strategy effectively reduced the model's variance and enhanced its stability and prediction accuracy. As shown on the right side of Table 2, the model's prediction errors (MAE, MSE, RMSE) increased significantly, and R² became negative. This is consistent with expectations, reflecting the model's limited generalization ability on heterogeneous populations with vastly different feature distributions and immune states. However, the key lies in the systematic bias presented by the model's prediction results: the predicted immune age of the vast majority of cancer patients was higher than their actual age. This trend aligns with the biological common sense of tumor immune aging, verifying from another perspective that the "immune age" signal captured by the model is related to the actual decline in immune system function. For individuals under 80 years old, this trend of "predicted age higher than actual age" is stable and significant, providing strong evidence for using this model to reveal immune aging under pathological conditions.

[0088] By combining the quantitative data in Tables 1 and 2, this invention demonstrates the superiority of its technical solution in the following aspects: 1. Through precise feature engineering and screening, excellent predictive performance is achieved with a very small number of core features; 2. Tree models (especially random forests optimized by Bagging) perform best in this task; 3. The model exhibits high accuracy and stability in healthy individuals and can reveal biologically expected immune aging trends in external samples (TM). This lays a solid performance foundation for the practical application of immune age assessment tools.

[0089] After describing the immune age prediction method of the present invention, the following is combined with... Figure 1 The process is described in further detail, along with corresponding device embodiments for implementing the method. Those skilled in the art will understand that the device can be implemented in hardware, software, or a combination of both. The immune age prediction device includes an acquisition module, a feature extraction module, and a prediction module connected in sequence.

[0090] The acquisition module is used to acquire blood sample test data of the individuals to be tested and construct an immune indicator dataset.

[0091] The feature extraction module is used to perform feature expansion processing on the immune index dataset using a deep feature synthesis algorithm to obtain an expanded feature set. Based on the pre-determined key features of immune age, the module selects a corresponding feature subset from the expanded feature set as the input features of the model.

[0092] The prediction module is used to input the input features into a pre-trained immune age prediction model to obtain the predicted immune age value of the individual to be tested.

[0093] The key features of immune age are predetermined through a process that includes the following steps: acquiring blood sample test data from multiple individuals to construct a training dataset on immune indicators; expanding the features of the training dataset using a deep feature synthesis algorithm; verifying the importance of features with respect to immune age using the Boruta algorithm based on the expanded feature set, and selecting a stable set of relevant features; and optimizing and selecting the stable set of relevant features using a forward selection algorithm, with the model's immune age prediction performance as a constraint, to obtain the set of key features with the fewest number of features while ensuring the predetermined prediction accuracy, as the key features of immune age.

[0094] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the immune age prediction method as described in 101-103 above.

[0095] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make modifications, alterations, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for selecting training set features for immune age prediction, characterized in that, Includes the following steps: Obtain immune indicator datasets from multiple individuals; use a deep feature synthesis algorithm to expand the features of the immune indicator datasets from multiple individuals; Based on the expanded feature set, the Boruta algorithm is used to verify the importance of features with respect to immune age, and a stable and relevant feature set is obtained by screening. By using a forward selection algorithm, with the model's immune age prediction performance as a constraint, a stable and relevant feature set is optimized and screened to obtain the minimum set of key features that ensures the predetermined prediction accuracy, which are then used as the key features for immune age to construct the training set.

2. The training set feature selection method for immune age prediction according to claim 1, characterized in that, The immune marker dataset includes immune cell subsets and mitochondrial function indicators obtained by flow cytometry, as well as routine hematological indicators obtained by blood analysis.

3. The training set feature selection method for immune age prediction according to claim 1, characterized in that, A deep feature synthesis algorithm is used to expand the features of immune indicator datasets from multiple individuals, including: In an immune indicator dataset belonging to an individual, four arithmetic primitive operations—addition, subtraction, multiplication, and division—are performed on immunological features to generate higher-order combined features in a recursive combination manner. Specifically, when performing multiplication and division arithmetic primitive operations, the immunological features are first logarithmized, and then addition is used to replace multiplication, and subtraction is used to replace division, in order to generate equivalent combined features.

4. The training set feature selection method for immune age prediction according to claim 1, characterized in that, Based on the expanded feature set, the Boruta algorithm was used to verify the importance of features with respect to immune age, and a stable set of relevant features was obtained, including: First, the mutual information algorithm is used to filter the expanded feature set to obtain a preliminary feature set that has a statistical dependence on age. Then, the Boruta algorithm is used to verify the immune age importance of the preliminary feature set and filter to obtain a stable and relevant feature set.

5. The training set feature selection method for immune age prediction according to claim 1, characterized in that, The forward selection algorithm optimizes and filters the stable relevant feature set under the constraint of the model's immune age prediction performance to obtain the key feature set with the smallest number of features while ensuring the predetermined prediction accuracy. This includes: gradually building a feature subset from an empty set through forward iteration; in each iteration, selecting a feature that can maximize the improvement of the immune age prediction model's performance from the stable relevant features that have not yet been selected into the feature subset; continuing the iteration until the model performance gain brought by the newly added feature is lower than the preset performance gain threshold; and determining the feature subset that has been built before the last iteration as the key feature set for immune age.

6. A method for predicting immune age, characterized in that, The immune index dataset of the individual to be tested is obtained. The immune index dataset is extended by using a deep feature synthesis algorithm to obtain an extended feature set. Based on the pre-determined key features of immune age, a corresponding feature subset is selected from the extended feature set as the model input features. The input features are input into the pre-trained immune age prediction model, and the corresponding immune age prediction value of the individual is obtained. The key features of immune age are determined using the training set feature selection method described in any one of claims 1-5; the immune age prediction model is trained based on the training set constructed from the key features of immune age.

7. The immune age prediction method according to claim 6, characterized in that, The pre-trained immune age prediction model is an ensemble model obtained by optimizing it using a Bagging ensemble learning strategy. The training process of the immune age prediction model includes: performing multiple bootstrap samplings with replacement on the training dataset to obtain multiple training subsets; and training a base prediction model independently using each training subset to obtain multiple base prediction models. The input features are fed into the immune age prediction model, which involves feeding them into multiple base prediction models. The immune age prediction values ​​output by the multiple base prediction models are then averaged to obtain the final immune age prediction value.

8. The immune age prediction method according to claim 6, characterized in that, The immune age prediction model is a machine learning model, which can be any one of the following: random forest model, lightweight gradient boosting tree model, extreme gradient boosting tree model, K-nearest neighbor model, and support vector machine model.

9. The immune age prediction method according to claim 6, characterized in that, The immune age prediction model is a tree-based machine learning model, which can be any one of the following: random forest model, lightweight gradient boosting tree model, and extreme gradient boosting tree model.

10. An immune age prediction device, characterized in that, include: The acquisition module is used to acquire the immune indicator dataset of the individual to be tested; The feature extraction module is used to perform feature expansion processing on the immune index dataset using a deep feature synthesis algorithm to obtain an expanded feature set. Based on the pre-determined key features of immune age, the module selects a corresponding feature subset from the expanded feature set as the input features of the model. The prediction module is used to input the input features into a pre-trained immune age prediction model and output the predicted immune age value for the corresponding individual. The key features of immune age are determined using the training set feature selection method described in any one of claims 1-5; the immune age prediction model is trained based on the training set constructed from the key features of immune age.

Images