A method for predicting intrinsic ability characteristics of the elderly based on machine learning

Abstract

This invention relates to the field of intrinsic ability prediction technology and discloses a method for predicting the intrinsic ability characteristics of the elderly based on machine learning. First, intrinsic ability score data of the elderly in five dimensions, including cognition and psychology, are acquired, and heterogeneity classification is performed using latent profile analysis. Using the classification results as labels, random forest and LASSO regression are used in parallel to screen key predictive factors to form a feature subset. A classification prediction model is built based on XGBoost, and hyperparameters are optimized through cross-validation and grid search. The SHAP and LIME algorithms are integrated to achieve both global and local interpretability. Finally, the model is deployed on an online interactive platform, where inputting individual characteristics outputs the IC propensity classification, probability, and feature contribution explanation. This invention achieves accurate classification and interpretable prediction of the intrinsic abilities of the elderly, reduces the dependence of assessment on professionals and the environment, and is suitable for large-scale application in grassroots elderly care scenarios.

Description

Technical Field

[0001] This invention relates to the field of intrinsic ability prediction technology, specifically a method for predicting the intrinsic ability characteristics of the elderly based on machine learning. Background Technology

[0002] Intrinsic capacity (IC) is considered to reflect changes in bodily function during the aging process. Compared to biological age, intrinsic capacity also takes into account cognitive and psychological aspects, and therefore has received much attention from scholars in the field of aging in recent years.

[0003] Currently, IC (Intensive Disorder) measurement requires a combination of multiple specialized scales, including MMSE (Mini-mental State Examination), GDS-15 (Geriatric Depression Scale), MNA-SF (Mini Nutritional Assessment Short Form), and SPPB (Short Physical Performance Battery). These scales are not only cumbersome to operate, requiring guidance from well-trained clinical professionals throughout the process, but also take 1 to 2 hours per test, significantly increasing the workload of healthcare workers. Furthermore, the test results are highly sensitive to the quietness and lighting conditions of the testing environment, as well as the patient's emotional fluctuations and cognitive alertness; even slight interference can lead to distorted assessment results. This severely limits their large-scale application in primary healthcare settings (such as community health service centers and township hospitals). Most importantly, for (semi)disabled (semi)demented elderly people, such scales rely heavily on verbal responses and physical cooperation. Some elderly people cannot accurately understand the test instructions due to cognitive impairment, or cannot complete the physical fitness test due to limited physical activity. This results in extremely poor applicability of the scales, which not only easily leads to misjudgment, but also increases the resistance of the elderly. Ultimately, this leads to low participation of respondents and makes it difficult to obtain comprehensive and accurate IC data.

[0004] Due to a severe shortage of healthcare personnel at the grassroots level, according to data from the National Health Commission, my country has fewer than 1.2 healthcare workers per 1,000 elderly people at the grassroots level, far below the standard of more than 2.5 in developed countries. Furthermore, these grassroots healthcare workers are burdened with multiple tasks, including daily diagnosis and treatment, chronic disease management, and health education, resulting in highly dispersed attention. Under these circumstances, using traditional scales for Integrated Clinical Assessment (IC) would consume a significant amount of healthcare resources. Therefore, it is urgent to rationally allocate limited healthcare resources (energy), focusing on the individual differences of the elderly and providing individualized, precise long-term care for the elderly, in order to improve the efficiency and quality of grassroots elderly care services.

[0005] Existing research primarily focuses on assessing the trajectory of cognitive function (IC), overemphasizing the dynamic fluctuations of individual IC over time while neglecting the heterogeneity of IC among different elderly populations (e.g., different sociodemographic characteristics, underlying diseases, and living environments). Furthermore, most studies only broadly examine the total IC score, ignoring the different contributions of dimensions such as cognitive function, emotional state, nutritional level, and physical fitness to the overall IC, thus failing to accurately pinpoint the weaknesses in the elderly's IC. Simultaneously, existing research is largely designed for hospital clinical settings, lacking adaptation to grassroots elderly care settings. It fails to consider the practical realities of limited equipment and the limited professional expertise of medical staff in grassroots settings, making it difficult to apply research findings practically. The assessment of IC lacks sufficient guidance for clinical elderly care, exhibiting problems such as weak assessment specificity, poor scenario adaptability, and inability to support precision care. Summary of the Invention

[0006] To address the aforementioned problems, the present invention aims to provide a machine learning-based method for predicting the intrinsic ability characteristics (IC) of older adults. This method effectively distinguishes the heterogeneity of IC among older adults, dynamically adapts to individual specificities while balancing real-time performance and universality, eliminates the need for full-time professional guidance, significantly shortens assessment time, avoids the high requirements of traditional scales on professional personnel and testing environments, and reduces environmental interference and human error. The technical solution is as follows:

[0007] A machine learning-based method for predicting intrinsic ability characteristics of older adults includes the following steps:

[0008] Step 1: Obtain internal ability assessment data from multiple elderly individuals. The internal ability assessment data shall include at least the score data for the cognitive dimension, mental health dimension, sensory function dimension, vitality dimension, and motor dimension.

[0009] Step 2: The latent profile analysis algorithm is used to perform cluster analysis on the intrinsic ability assessment data obtained in Step 1. Based on the model fitting and comparison index, the optimal cluster category of intrinsic ability is determined, and the heterogeneity classification results of the intrinsic ability of each elderly person are obtained.

[0010] Step 3: Use the heterogeneity classification results of intrinsic ability obtained in Step 2 as label variables, and use the relevant characteristics of heterogeneity of intrinsic ability of the elderly as candidate predictors. Use three algorithms in parallel, namely univariate regression, random forest algorithm and LASSO regression, to screen features, extract key predictors, and form the final feature subset.

[0011] Step 4: Using the final feature subset obtained in Step 3 as input features and the intrinsic ability heterogeneity classification result obtained in Step 2 as the prediction target, a classification prediction model is constructed using the XGBoost gradient boosting algorithm. The classification prediction model is trained using the training dataset, and the model hyperparameters are optimized through cross-validation and grid search.

[0012] Step 5: Use the Shapley additive interpretation algorithm to perform a global interpretation on the trained classification prediction model, quantifying the average contribution of each input feature to the model's prediction results; and use the locally interpretable model-independent interpretation algorithm to perform a local interpretation on the classification prediction model, quantifying the importance contribution of each input feature to the prediction results of a single sample.

[0013] Step 6: Deploy the trained and interpretability-enhanced classification prediction model to the online interactive platform, obtain multi-dimensional feature data of the target elderly person and input it into the classification prediction model, output the prediction results and corresponding probabilities of the elderly person belonging to each intrinsic ability heterogeneity subtype, and simultaneously output the feature contribution interpretation information generated by the Shapley additive interpretation algorithm and / or the locally interpretable model-independent interpretation algorithm.

[0014] The beneficial effects of this invention are:

[0015] 1) Addressing the prominent issues of time-consuming, labor-intensive, and highly volatile IC assessments in primary healthcare settings, this invention achieves efficient and accurate assessment through an innovative technological approach. Traditional assessments rely on multiple scale combinations, which are cumbersome and time-consuming. In contrast, this invention uses core data from the five dimensions of IC to perform clustering and typing through LPA (Latent Profile Analysis). It combines random forest and LASSO (Least Absolute Shrinkage and Selection Operator) regression to screen key predictive factors, constructs an XGBoost prediction model, and deploys an online interactive platform. No professional guidance is required throughout the process; users can quickly obtain IC propensity prediction results by inputting relevant features. This significantly shortens assessment time, avoids the high requirements of traditional scales on professional personnel and testing environments, reduces environmental interference and human error, effectively solves the practical dilemma of insufficient and scattered medical staff at the grassroots level, reduces the workload of medical staff, and is suitable for large-scale application needs in primary healthcare settings.

[0016] 2) Addressing the shortcomings of existing technologies in neglecting individual heterogeneity of intracranial pressure (IC) and poor application adaptability, this invention achieves breakthroughs in precision and scenario-based application. Existing technologies often focus on the dynamic trajectory of IC and broadly assess the overall score. In contrast, this invention uses LPA to precisely classify IC clusters and combines SHAP (Shapley Additive Explanations) and LIME (Local Interpretable Model-agnostic Explanations) algorithms to clearly quantify the contribution weight of each dimension to IC classification, accurately pinpointing the weak points of IC in different elderly populations and fully highlighting the individual heterogeneity of IC. Furthermore, this invention is grounded in grassroots elderly care scenarios. The model construction takes into account the limited equipment and professional expertise available at the grassroots level. The online interactive platform is user-friendly and requires no complex equipment, improving the relevance and reliability of IC assessment while breaking down the barriers between technological achievements and grassroots applications. This allows IC assessment to truly serve individualized and precise elderly care, compensating for the shortcomings of existing technologies in terms of application value. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method for predicting the intrinsic ability characteristics of the elderly based on machine learning, according to the present invention.

[0018] Figure 2 LPA classification for intrinsic capabilities.

[0019] Figure 3 This represents the average descent accuracy of each variable in the random forest model.

[0020] Figure 4 This represents the average decrease in the Gini coefficient of each variable in the random forest model.

[0021] Figure 5 The binomial deviation curve based on the negative log-likelihood ratio (LASSO regularization parameter) Select image).

[0022] Figure 6 Rank the variables by importance.

[0023] Figure 7 The results of the evaluation of the XGBoost model estimation index are presented.

[0024] Figure 8 SHAP plot for each variable.

[0025] Figure 9 This is a LIME chart using NO.24 as an example. Detailed Implementation

[0026] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0027] This invention utilizes population survey data on IC (intellectual, mental health, sensory function, vitality, and motor skills) across five dimensions. It employs Mplus 8.3 software to conduct Latent Profile Analysis (LPA), classifying IC levels using these five dimensions as core indicators. To determine the optimal latent category model, the Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), adjusted Bayesian Information Criterion (aBIC), Lomendel-Rubin test (LMRT), Bootstrap Likelihood Ratio Test (BLRT), and entropy are selected as model fitting and comparison indicators. AIC, BIC, and aBIC are used to evaluate the model fit. The smaller the value, the better the trade-off between model fit and complexity. When the p-value of LMRT and BLRT is less than 0.05, it indicates that the k-norm model fits better than the k-1 norm model. The entropy value ranges from 0 to 1, and ≥0.80 indicates that the model classification quality is good. Finally, the optimal clustering category of IC is determined by combining the above indicators.

[0028] Based on the IC clustering classification results obtained from LPA, the correlation between IC and demographic indicators such as age, gender, and education level was further analyzed. Random forest and LASSO regression models were used for feature selection to identify key predictive factors significantly associated with IC classification. In the random forest model, continuous variables were automatically split into nodes by searching for the optimal splitting threshold, while categorical variables were directly included as factors. The importance of features was ranked by calculating the average Gini index decrease during the splitting process of each feature node. In the L1-regularized LASSO regression model, the contribution of features was ranked according to the size of the variable coefficients, and L1 regularization was used to bring the coefficients of variables with smaller contributions closer to 0. Finally, combining the selection results of the two models, variables that showed high contribution in both models were selected as the final feature subset for constructing the IC classification prediction model.

[0029] Using key variables selected from univariate regression, random forest, and LASSO regression models as input features, an IC classification prediction model was constructed based on the XGBoost gradient boosting algorithm. First, the dataset was randomly divided into a training set (70%) and an independent test set (30%) in a 7:3 ratio. The training set was used for feature selection and model development, while the independent test set was used to evaluate model effectiveness. On the training set, five-fold cross-validation combined with grid search was used to optimize the model's hyperparameters. The hyperparameters of the fold with the best performance in cross-validation were selected for final model training. Accuracy was used as the evaluation metric, and the model's generalization ability was further tested on the independent test set. Simultaneously, the model performance was evaluated from four dimensions: discriminative power, clinical applicability, calibration, and generalization ability. After model training, the XGBoost prediction model was deployed to an online interactive portal. Users can input multi-dimensional individual feature information to automatically obtain the prediction results and corresponding probabilities of heterogeneous IC tendency, enabling rapid identification and assisted judgment.

[0030] To enhance the interpretability and credibility of the XGBoost IC classification prediction model, the Shapley Additive Interpretation (SHAP) algorithm and the Locally Interpretable Model-Independent Interpretation (LIME) algorithm were employed to quantify the contribution of features to the model. The SHAP algorithm takes a global perspective, calculating the Shapley value of each feature within a feature combination to evaluate its average contribution to the model, and simultaneously analyzing the impact of each feature on the overall prediction result by calculating the weighted average of the Shapley values. The LIME algorithm, on the other hand, takes a local perspective, quantifying the importance of each feature to the prediction result of a single sample. By combining the interpretation results of both algorithms, key features that significantly influence IC classification were clearly identified, the model's prediction logic was clearly explained, and the persuasiveness and clinical application value of the model results were enhanced.

[0031] The specific implementation steps are as follows:

[0032] Step 1: Data acquisition and preprocessing.

[0033] Based on the five dimensions of the cognitive ability (IC) obtained from a population survey, intrinsic ability assessment data were acquired for multiple elderly individuals. This intrinsic ability assessment data included at least scores for the cognitive, mental health, sensory function, vitality, and motor dimensions. The intrinsic ability assessment data underwent preprocessing such as cleaning and filtering.

[0034] Step 2: Classification of intrinsic ability heterogeneity.

[0035] The latent profile analysis (LPA) algorithm was used to perform cluster analysis on the intrinsic ability assessment data obtained in step 1. Based on model fitting and comparison indicators, the optimal cluster category of intrinsic ability was determined, and the heterogeneity classification results of intrinsic ability of each elderly person were obtained.

[0036] For the five dimensions of intrinsic ability assessment data of the elderly obtained in Step 1, Latent Profile Analysis (LPA) clustering was used. The core difference between LPA and traditional clustering (such as K-means) is that LPA does not require a pre-set number of clusters; it can objectively determine the optimal category through fitting indices, avoiding subjective bias and better reflecting the heterogeneity of intrinsic ability among the elderly. The data was first preprocessed (missing values, outliers, and dimensionality handling), and then n (n≥2) LPA models were fitted step-by-step, with the optimal category selected through multiple indices. Among the core indices, the entropy value ranged from 0 to 1; the closer to 1, the higher the classification accuracy and discrimination; values ​​below 0.7 required refitting. Smaller AIC (Akaike Information Content) and BIC (Akaike Information Content) values ​​resulted in better model fitting and avoided overfitting. The Lomendel-Rubin test (LMRT) and the bootstrap likelihood ratio test (BLRT) are used to examine the goodness of fit of the models. Both compare the goodness of fit between the k-class and k-1-class models. The null hypothesis is that "the k-1-class model is sufficiently well-fitted." If the p-value is <0.05, the null hypothesis is rejected, indicating that the k-class model is significantly better than the k-1-class model, and more classes need to be added. If the p-value is ≥0.05, the null hypothesis is accepted, and no more classes need to be added; the optimal number of classes is k-1. By combining the above indicators with theoretical and clinical significance, the optimal classes are determined, the characteristics of each subtype are clarified, and the results of the heterogeneity classification of intrinsic abilities in the elderly are obtained.

[0037] The core fit index formula for Potential Profile Analysis (LPA) is as follows:

[0038] (1) Entropy (classification accuracy / discrimination index)

[0039] The entropy value ranges from [0,1]. The closer it is to 1, the higher the classification accuracy and discrimination. The formula is as follows:

[0040] ;

[0041] in: The total sample size (total number of elderly people). The number of classes fitted to the LPA model (K≥2); Let be the posterior probability (output by the LPA model) of the i-th sample being assigned to the k-th class; if ,but (To avoid logarithmic inconsistency).

[0042] (2) Akaike Information Criterion (AIC, model fit index)

[0043] Akaike Information Criterion Value The smaller the value, the better the model fit, which can avoid overfitting. The formula is as follows:

[0044] ;

[0045] in: This represents the total number of parameters to be estimated in the LPA model (including prior probabilities for each category, distribution parameters of each dimension under each category, etc.). This is the log-likelihood value of the model (reflecting how well the model fits the data; a larger value indicates a better fit). The P-value is calculated based on the corrected asymptotic chi-square distribution.

[0046] (3) Bayesian Information Criterion (BIC, model fit index)

[0047] BIC places greater emphasis on avoiding overfitting than AIC, and the Bayesian information criterion value... A smaller value indicates a better fit, as shown in the formula below:

[0048] ;

[0049] in: This represents the total number of parameters to be estimated in the model (as defined in the AIC formula). The total sample size (as defined in the entropy formula); This is the log-likelihood value of the model (as defined in the AIC formula). The p-value is calculated based on the empirical distribution of 1000 parameter bootstrapping samples.

[0050] (4) Likelihood ratio-based test, namely Lomendel-Rubin test (LMRT, number of categories selection test)

[0051] The Lomendel-Rubin test is used to examine the difference in goodness of fit between the k-type model and the k-1-type model. The null hypothesis H0 is that the k-type model and the k-1-type model have no difference in goodness of fit; the alternative hypothesis H1 is that the k-type model is better than the k-1-type model. The test statistic is the Lomendel-Rubin likelihood ratio. The calculation formula is as follows:

[0052] ;

[0053] in: The log-likelihood value is for the k-1 class model; Let be the log-likelihood value of the k-type model.

[0054] The model follows a chi-square distribution, with the degrees of freedom being the difference in the number of parameters between the two models. If P < 0.05, the null hypothesis is rejected, indicating that class k is significantly better than class k-1, and more classes need to be added.

[0055] (5) Self-help likelihood ratio test (BLRT, number of categories selection test)

[0056] The bootstrap likelihood ratio test functions similarly to the Lomendel-Rubin test, but is more suitable for small samples. The test statistic is the same as the Lomendel-Rubin test; the core difference lies in the calculation method of the p-value (obtained through bootstrap sampling). The test statistic is the bootstrap likelihood ratio. The formula is as follows:

[0057] ;

[0058] in: , The definition is the same as the LMRT formula.

[0059] By performing multiple Bootstrap samplings, the distribution of the test statistic under the null hypothesis is simulated, and then the P-value is calculated. The judgment criteria are the same as LMRT (P<0.05 requires adding a category, P≥0.05 does not require adding a category).

[0060] Step 3: Screening of key predictors.

[0061] The intrinsic ability heterogeneity classification results obtained in step 2 are used as label variables. The intrinsic ability heterogeneity-related characteristics of the elderly are used as candidate predictors. Three algorithms, namely univariate regression, random forest and LASSO regression, are used in parallel to screen features and accurately extract key predictors to form the final feature subset.

[0062] Each algorithm focuses on clearly defined importance metrics and judgment criteria: In univariate regression analysis, the key considerations are P-value (statistical significance), hazard ratio Exp(B) (effect direction and strength), and the corresponding 95% confidence interval (to judge the stability of the results); in the random forest algorithm, the key focus is on the MDA value (Mean Decrease Accuracy, reflecting the contribution of a feature to the model's prediction); in the LASSO regression algorithm, the key consideration is the coef. value (regression coefficient; a non-zero coefficient or a criterion-compliant absolute value indicates that the feature has predictive value). Finally, strictly adhering to the intersection selection principle, candidate predictive factors that simultaneously meet the preset importance conditions in both the random forest and LASSO regression algorithms are selected and merged to form the final feature subset, providing scientific feature support for subsequent construction of intrinsic capability heterogeneity prediction models and the formulation of precise intervention strategies.

[0063] Feature importance assessment (or feature importance evaluation) is used to calculate the importance of sample features and quantitatively describe the contribution of features to classification or regression. Random forests can be used to evaluate feature importance. From another perspective, feature importance assessment is a built-in tool of random forests, mainly divided into two methods: (1) Mean Decrease Impurity (MDI) measures the importance of a node by statistically measuring the decrease in impurity when a node splits; (2) Mean Decrease Precision (MDA) method randomly swaps the value of a feature in the permuted oob data and then re-predicts, calculating the importance of the feature by measuring the degree of decrease in classification / regression accuracy.

[0064] (1) Classification

[0065] ;

[0066] Where T is the number of random trees in RF (Random Forest); For a sample, Y represents the category for classification and the output (which can be multidimensional) for regression. For the sample The sample after random swapping of the j-th dimension (feature); Let be the out-of-bag (oob) sample set of the random tree t; The sample set formed after the j-th dimension is swapped; For the sample The predicted results (categories); This is an indicator function that returns 1 if the prediction matches the true class, and 0 otherwise.

[0067] (2) LASSO regression core loss function

[0068] LASSO regression compresses regression coefficients through L1 regularization (L1 penalty) to filter out features with non-zero coefficients. The core loss function (taking linear regression as an example) is:

[0069] ;

[0070] If the label variable is categorical (logistic LASSO regression), the loss function is replaced with log-likelihood loss:

[0071] ;

[0072] in, For the total number of elderly people, The number of candidate predictors; For the first The label variable values ​​for each sample; For the first The first sample The values ​​of the candidate predictors; For the intercept term, For the first The regression coefficients (coef. values) of each candidate predictor. The regularization parameter (penalty coefficient) is used to determine the optimal value through cross-validation (such as 10-fold cross-validation). The coefficient vector is a set of regression coefficients for all candidate predictors, containing ( , , ..., )common Each dimension corresponds to The effect size of each predictor.

[0073] The core focus is on the measurement value (coef. value) and the judgment criteria.

[0074] The coef value is obtained after solving the LASSO regression. (Regression coefficient estimates), judgment criteria:

[0075] Basic standards: This indicates that the predictor was retained by LASSO regression screening and has predictive value;

[0076] Advanced standards: Set absolute value thresholds for coefficients based on research needs (e.g., , (As a preset minimum value), select predictive factors with higher influence intensity.

[0077] Step 4: Prediction model construction and training.

[0078] Using the final feature subset obtained in step 3 as input features and the intrinsic ability heterogeneity classification result obtained in step 2 as the prediction target, a classification prediction model is constructed using the XGBoost gradient boosting algorithm. The classification prediction model is trained using the training dataset, and the model hyperparameters are optimized through cross-validation and grid search.

[0079] Step 5: Enhance model interpretability.

[0080] The Shapley additive interpretation algorithm is used to provide a global interpretation of the trained classification prediction model, quantifying the average contribution of each input feature to the model's prediction results. The locally interpretable model-independent interpretation algorithm is used to provide a local interpretation of the classification prediction model, quantifying the importance contribution of each input feature to the prediction results of a single sample.

[0081] After training the classification prediction model (step 4), a dual interpretation is performed using the Shapley Additive Interpretation (SHAP) algorithm and the Locally Interpretable Model-Independent Interpretation (LIME) algorithm. The SHAP algorithm is used for global interpretation, quantifying the average contribution of each input feature to the overall prediction result of the model and clarifying the overall pattern of feature influence. The LIME algorithm is used for local interpretation, accurately quantifying the importance contribution of each input feature to the prediction result of a single sample and analyzing individual prediction differences. The core innovation of this scheme lies in constructing a dual interpretation mechanism of SHAP and LIME, combining real-time SHAP / LIME interpretation feedback to break the limitations of the traditional "black box" model, achieving bidirectional analysis of the global patterns and individual differences in the prediction results, and improving the credibility and practicality of the model's interpretation.

[0082] 1. SHAP Algorithm Formula

[0083] The Shapley value of each feature in the feature combination is calculated using the Shapley additive interpretation algorithm to evaluate its average contribution to the model. The impact of each feature on the overall prediction results is analyzed by calculating the weighted average of the Shapley values ​​of each feature.

[0084] (1) Core Additive Explanation Model

[0085] Using an interpretable additive model To approximate the complex black box model to be explained Prediction results:

[0086] ;

[0087] in: For the explanation model (interpretable additive model); The union vector indicates whether the feature exists; The total number of features; The baseline value (usually the average predicted value of all samples) ); Features The SHAP value represents the contribution of the feature to the prediction result. For coalition vectors The The nth component represents the nth component. The existence of a feature is determined by whether it has a value of 0 or 1.

[0088] (2) Standard SHAP value formula (based on Shapley value)

[0089] For complex black box models and input samples Define features The standard SHAP value, i.e., the Shapley weighted average. for:

[0090] ;

[0091] in: For the set of all features, ; For features not included A subset of features; For subset Size (number of features); This is a background distribution mapping function used to generate a subset-only distribution. Samples with features; Features Join a subset Marginal contribution after; As a weighting factor, it ensures that all permutations and combinations of features are considered fairly.

[0092] 2. LIME (Locally Interpretable Model-Independent Interpretation) Algorithm Formula:

[0093] (1) Core optimization objective function

[0094] LIME interprets prediction results by constructing an interpretable surrogate model in the local neighborhood of the target sample. Its optimization objective is:

[0095] ;

[0096] in, A complex black box model to be explained; For interpretable surrogate models (such as sparse linear models); For the set of all possible interpretable models; The local fidelity loss function measures... and Approximation to the target sample; This is a penalty term for model complexity, ensuring that the explanation is easy to understand.

[0097] (2) Loss function and weight calculation

[0098] 1) Weighted loss function

[0099] ;

[0100] in, In the target sample A collection of synthetic samples generated nearby; For synthetic samples Interpretable feature representation, also known as simplified feature representation; These are sample weights, reflecting the similarity to the target sample; Original black box model For synthetic samples The predicted output.

[0101] 2) Commonly used weight kernel functions

[0102] LIME typically uses an exponential kernel to calculate sample weights:

[0103] ;

[0104] in, For target samples With synthetic samples The distance between them (e.g., Euclidean distance); The kernel width parameter controls the local neighborhood range.

[0105] Step 6: Model Deployment and Application.

[0106] The trained and interpretability-enhanced classification prediction model is deployed to an online interactive platform. Multi-dimensional feature data of the target elderly person is obtained and input into the classification prediction model. The model outputs the prediction results and corresponding probabilities of the elderly person belonging to each intrinsic ability heterogeneity subtype. Simultaneously, the feature contribution explanation information generated by the Shapley additive interpretation algorithm and / or the locally interpretable model-independent interpretation algorithm is output.

[0107] Case Study:

[0108] (1) Taking 4,508 elderly people from 24 provinces and cities across the country as an example, their scores in the five dimensions of intrinsic ability are shown in Appendix Table 1. Based on this, latent category models 1 to 3 were constructed. The results show that the Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), and adjusted BIC (aBIC) values ​​of models 1 to 3 show a decreasing trend. The P values ​​of the Lomendel-Rubin Test (LMRT) and Self-Help Likelihood Ratio Test (BLRT) of models 1 to 2 are all below 0.05. However, the LMRp value of model 3 is close to 0.05 (P=0.588). Following the principle of conservative and cautious model fitting, we found that model 2 shows a higher classification accuracy (entropy = 0.846, P < 0.001). Therefore, the respondents' IC levels were divided into two potential groups and named according to their characteristics across five dimensions: class 1: low physical-high psychological type; class 2: high physical-low psychological type (hereinafter referred to as high physical-low psychological IC), as detailed in Table 1 and... Figure 2 .

[0109] Table 1. LPA Classification of Intrinsic Competencies

[0110] .

[0111] (2) The results of univariate analysis and random forest analysis on the tendency of high somatic-low psychological IC are shown in Table 2. Figure 3 and Figure 4 As shown, the LASSO regression results are as follows: Figure 5 See Table 3. At this stage of the analysis, 15 variables were determined to be statistically significant (P < 0.05). In the random forest analysis, after testing and tuning, the estimation error rate for out-of-bag samples was minimized at ntree = 610 and mtry = 3, resulting in an out-of-bag error rate of 21.23%. The five most important feature variables identified in the random forest analysis are... Figure 3The study included self-rated health scores, (other-rated) health status, mobility impairments (slow walking / limping), age, and education level. Notably, univariate analysis showed a negative correlation between age and high somatic-low psychogenic IC. This result suggests that the likelihood of a high somatic-low psychogenic IC tendency decreases with age. Furthermore, individuals reporting high self-rated health scores (OR = 1.62, 95% CI = 1.17–1.36) or multiple chronic diseases (OR = 1.146, 95% CI = 1.054–1.247), chest tightness / chest pain (OR = 1.283, 95% CI = 1.016–1.619), and mobility impairments (slow walking / limping) (OR = 2.541, 95% CI = 2.076–3.109) were more likely to have a high somatic-low psychogenic IC tendency. For the training set, independent predictors of relapse tendency were identified using LASSO regression. Finally, by integrating the results of these two analytical methods, 11 best predictors of relapse tendency were identified: self-rated health score, age, peer-rated health score, education level, average monthly income per person, mobility problems (slowness / claudication), number of chronic diseases, major occupations throughout life, marital status, economic status, and dizziness. The specific results are shown in Figure 6.

[0112] Table 2. Selection of key variables using univariate regression and random forest models.

[0113] .

[0114] Table 3. Estimation of Lambda Regularization Parameters in LASSO Regression

[0115] .

[0116] (3) Based on the confusion matrix classification model performance evaluation tool, the results present the metrics used to evaluate the model's classification performance. See Figure 7 In the test set, although the AUC (Area Under the Curve) and accuracy of the XGBoost model were not higher than 0.85, they were both higher than 0.80 (AUC = 0.851; accuracy = 0.835). Its recall and F1 index were higher than 0.9 (Recall = 0.946; F1 = 0.888), but its specificity was poor (Specificity = 0.369). These results indicate that the model has strong discriminative ability and excellent diagnostic efficacy in the independent test set.

[0117] (4) Subsequently, the XGBoost Model was interpreted using the SHAP algorithm. Based on the interpretation results, the influence and relative importance ranking of the risk factors included in the model on the occurrence of the high physical-low psychological IC propensity IC classification were obtained. The importance ranking of risk factors was based on the sum of the absolute values ​​of the SHAP values ​​corresponding to each feature of all subjects in the test set with high physical-low psychological IC propensity classification, ranked from high to low. The importance of each risk factor was obtained under the interaction with other risk factors. Figure 8 The study presents the 11 most important risk factors for predicting a high somatic-low psychotic IC propensity, based on SHAP analysis. The factors, ranked from most to least important, are: self-rated health score, age, health status, education level, average monthly income, mobility impairment, number of chronic diseases, main occupation, marital status, economic situation, and dizziness. Taking respondent NO.24 as an example, this SHAP diagram visually presents the driving and inhibiting effects of various characteristics on the model's prediction results. The overall prediction baseline E[f(x)] is 1.57, and the final predicted value f(x) for this sample reaches 3.84, far exceeding the baseline, indicating that this respondent is more likely to exhibit a high somatic-low psychological IC tendency. Among them, the self-rated health score (9 points) is the most crucial factor that raises the predicted value. Poor other-rated health status, average monthly disposable income (3000-5999 yuan), no slowness / limping problems, no dizziness symptoms, a lifelong primary occupation mainly consisting of light physical labor, and marital status of being married and living together all have a positive driving effect on the predicted value. The age of 79 years old is the most significant inhibiting factor, and the education level (primary school) also lowers the predicted result with a contribution of -0.294. Figure 9 As shown.

[0118] (5) Deploy the trained XGBoost IC classification prediction model on an online interactive prediction portal.

[0119] This system is compiled using Python 3.9, JavaScript 7, and MySQL 8.0, and can be developed using cross-platform software such as Visual Studio Code. After installing the compilation environment, it can be deployed using a Docker container and accessed via a web page.

[0120] This invention achieves innovative precision classification and feature mining: for the first time, it combines Latent Profile Analysis (LPA) with a dual feature selection model to accurately classify the intrinsic abilities (IC) clusters of older adults. By leveraging the complementary advantages of random forests and LASSO regression, it efficiently identifies core predictive factors highly correlated with IC classification, overcoming the limitations of single-model selection and improving the scientific rigor and reliability of feature selection.

[0121] This invention enables interpretable intelligent prediction and application innovation: It constructs and deploys an IC classification and prediction model based on XGBoost, breaking through the "black box" bottleneck of traditional machine learning. It innovatively integrates SHAP and LIME algorithms to achieve global and local dual-dimensional interpretation, while simultaneously building an online interactive portal website to enable rapid, accurate identification and accessible application of high physical-low psychological IC propensity in the elderly.

Claims

1. A method for predicting the intrinsic ability characteristics of older adults based on machine learning, characterized in that, Includes the following steps: Step 1: Obtain internal ability assessment data from multiple elderly individuals. The internal ability assessment data shall include at least the score data for the cognitive dimension, mental health dimension, sensory function dimension, vitality dimension, and motor dimension. Step 2: The latent profile analysis algorithm is used to perform cluster analysis on the intrinsic ability assessment data obtained in Step 1. Based on the model fitting and comparison index, the optimal cluster category of intrinsic ability is determined, and the heterogeneity classification results of the intrinsic ability of each elderly person are obtained. Step 3: Use the heterogeneity classification results of intrinsic ability obtained in Step 2 as label variables, and use the relevant characteristics of heterogeneity of intrinsic ability of the elderly as candidate predictors. Use three algorithms in parallel, namely univariate regression, random forest algorithm and LASSO regression, to screen features, extract key predictors, and form the final feature subset. Step 4: Using the final feature subset obtained in Step 3 as input features and the intrinsic ability heterogeneity classification result obtained in Step 2 as the prediction target, a classification prediction model is constructed using the XGBoost gradient boosting algorithm. The classification prediction model is trained using the training dataset, and the model hyperparameters are optimized through cross-validation and grid search. Step 5: Use the Shapley additive interpretation algorithm to perform a global interpretation on the trained classification prediction model, quantifying the average contribution of each input feature to the model's prediction results; and use the locally interpretable model-independent interpretation algorithm to perform a local interpretation on the classification prediction model, quantifying the importance contribution of each input feature to the prediction results of a single sample. Step 6: Deploy the trained and interpretability-enhanced classification prediction model to the online interactive platform, obtain multi-dimensional feature data of the target elderly person and input it into the classification prediction model, output the prediction results and corresponding probabilities of the elderly person belonging to each intrinsic ability heterogeneity subtype, and simultaneously output the feature contribution interpretation information generated by the Shapley additive interpretation algorithm and / or the locally interpretable model-independent interpretation algorithm.

2. The method for predicting the intrinsic ability characteristics of the elderly based on machine learning according to claim 1, characterized in that, The model fitting and comparison metrics used in step 2 include: entropy, Akaike information criterion, Bayesian information criterion, adjusted Bayesian information criterion, Lomendel-Rubin likelihood ratio test, and bootstrap likelihood ratio test. Entropy The value ranges from [0,1]. The closer it is to 1, the higher the classification accuracy and discrimination. The formula is as follows: ； in, The total number of elderly people; The number of classes fitted to the potential profile analysis algorithm model, and ≥2; the starting point of the inner summation must be k=1, and the ending point is ; The i-th sample is assigned to the i-th group The posterior probability of each category; Akaike Information Criterion Value The smaller the value, the better the model fit. The formula is as follows: ； in, and These represent the total number of parameters to be estimated and the log-likelihood value of the potential profile analysis algorithm model, respectively. Bayesian Information Criterion Values The smaller the value, the better the model fit. The formula is as follows: ； Using Lomendel-Rubin likelihood ratio The test statistic is given by the following formula: ； in, This represents the maximum likelihood estimate of the k-1 category potential profile model. The maximum likelihood estimate of the potential profile model for category k is given; the P-value is calculated based on the corrected asymptotic chi-square distribution. Self-help likelihood ratio The formula is as follows: ； It is based on N s The p-value is calculated based on the empirical distribution of the secondary parameter auto-sampling. The model with a p-value less than 0.05 and an entropy value greater than or equal to 0.80 in the Lomendel-Rubin test and the bootstrap likelihood ratio test is considered the optimal clustering class.

3. The method for predicting the intrinsic ability characteristics of the elderly based on machine learning according to claim 1, characterized in that, In step 3, the random forest algorithm uses average reduction purity or average reduction precision as a measure of feature importance. The average reduction in precision is calculated as follows: ； Where T is the number of random trees in the random forest; For a sample, Y represents the category for classification and the output for regression. For the sample The sample after randomly swapping the j-th dimension feature; Let t be the out-of-bag sample set of the random tree t; The sample set formed after the j-th dimension is swapped; For the sample Category prediction results; This is an indicator function that returns 1 if the prediction matches the true class, and 0 otherwise.

4. The method for predicting the intrinsic ability characteristics of the elderly based on machine learning according to claim 1, characterized in that, In step 3, LASSO regression compresses the regression coefficients using L1 regularization, filtering out features with non-zero coefficients. The core loss function is: ； If the label variable is categorical, the loss function is replaced with log-likelihood loss: ； in, For the total number of elderly people, The number of candidate predictors; For the first The label variable values ​​for each sample; For the first The first sample The values ​​of the candidate predictors; For the intercept term, For the first Regression coefficients of candidate predictors; The regularization parameter is determined through cross-validation to find the optimal value. The coefficient vector consists of the regression coefficients of all candidate predictors, containing , , ..., ,common Each dimension corresponds to The effect size of each predictor; The regression coefficient estimates obtained after solving the LASSO regression problem. The criteria for judgment are: Basic standards: If If the current candidate predictor is selected and retained by LASSO regression, it has predictive value; Advanced criteria: Set a threshold for the absolute value of the coefficient to filter out predictive factors with higher influence.

5. The method for predicting the intrinsic ability characteristics of the elderly based on machine learning according to claim 1, characterized in that, In step 3, the final feature subset includes at least one of the following features: self-rated health score, age, peer-rated health score, education level, average monthly income per person, mobility problems, number of chronic diseases, major occupations throughout life, marital status, economic status, and dizziness.

6. The method for predicting the intrinsic ability characteristics of the elderly based on machine learning according to claim 1, characterized in that, Step 5 specifically includes: Step 5.1: Calculate the Shapley value of each feature in the feature combination using the Shapley additive interpretation algorithm, evaluate its average contribution to the model, and analyze the impact of each feature on the overall prediction result by calculating the weighted Shapley average of each feature value; specifically: Using an interpretable additive model To approximate the complex black box model to be explained Prediction results: ； in, The union vector indicates whether the feature exists; The total number of features; As the baseline value; Features The SHAP value represents the contribution of the feature to the prediction result; For coalition vectors The The nth component represents the nth component. The existence of a feature is indicated by a value of 0 or 1. For the complex black box model and input samples Define features The standard SHAP value, i.e., the Shapley weighted average. for: ； in, The set of all features, the number of features ; For features not included A subset of features; For subset The number of features; This is a background distribution mapping function used to generate a subset-only distribution. Samples with features; Features Join a subset Marginal contribution after; As a weighting factor, it ensures that all permutations and combinations of features are considered fairly; Step 5.2: The locally interpretable model-independent interpretation algorithm constructs an interpretable surrogate model in the local neighborhood of the target sample, quantifying the importance of each feature to the prediction result of a single sample. Its optimization objective is: ； Among them, interpretable additive models This is a locally fitted model; For the set of all possible interpretable models; For target samples For the local neighborhood, define an interpretable additive model. The fitting range; The local fidelity loss function measures... and In the local neighborhood of the target sample The degree of approximation within; This serves as a penalty for model complexity, ensuring that the explanation is easy to understand. The local fidelity loss function is calculated as follows: ； in, In the target sample A collection of synthetic samples generated nearby; For synthetic samples Interpretable feature representation; These are sample weights, reflecting the similarity to the target sample; For complex black box models For synthetic samples The predicted output; The sample weight The calculation is as follows: ； in, For target samples With synthetic samples The distance between them; This is the kernel width parameter, used to control the local neighborhood range.

Images