基于长短期记忆网络对中老年人无残疾预期寿命轨迹预测模型的构建

Construction of a Trajectory Prediction Model for Disability-Free Life Expectancy in Middle-Aged and Elderly People Based on Long Short-Term Memory Networks

摘要: 目的：构建并验证中老年人无残疾预期寿命的轨迹预测模型。方法：基于CHARLS数据库的面板数据，选择≥45岁具有完整随访资料的中老年人作为研究对象，按照7:3比例随机分为构建集(n = 7826)和验证集(n = 3354)。使用Python进行数据清洗以及特征处理，采用LSTM模型构建无残疾预期寿命的轨迹预测模型。使用SHAP值表示模型预测结果的贡献度,通过ROC曲线下面积确定模型的拟合优度和预测效果，并以验证集进行外部验证。结果：残疾状态的发生率从2011年的16%上升至2020年的24%。构建集模型ROC曲线下面积为0.788 (95% CI: 0.603~0.798)，敏感度为81.3%，特异度为86.2%，校准曲线与理想曲线相近，Brier得分为0.115；验证集模型ROC曲线下面积为0.745 (95% CI: 0.668~0.865)，敏感度83.9%，特异度为85.5%。SHAP值显示影响中老年人残疾状态的主要因素包括年龄、慢性病数量、关节炎、睡眠时间和性别等。结论：本研究构建的轨迹预测模型能够较好地预测中老年人无残疾预期寿命，可以为早期预防和护理决策提供支持。

Abstract: OBJECTIVE: To construct and validate a predictive model for the trajectory of disability-free life expectancy in middle-aged and elderly populations. METHODS: Utilizing panel data from the China Health and Retirement Longitudinal Study (CHARLS) database, individuals aged ≥ 45 years with complete follow-up records were included as study participants. The cohort was randomly divided into a training set (n = 7826) and a validation set (n = 3354) in a 7:3 ratio. Data preprocessing, including data cleaning and feature engineering, was conducted using Python. A Long Short-Term Memory (LSTM) model was employed to construct a trajectory prediction model for disability-free life expectancy. The SHapley Additive exPlanations (SHAP) values were utilized to assess feature contributions to the model’s predictions. Model performance, including goodness-of-fit and predictive accuracy, was evaluated using the area under the receiver operating characteristic (ROC) curve (AUC), and external validation was conducted using the validation dataset. RESULTS: The prevalence of disability increased from 16% in 2011 to 24% in 2020. The AUC for the training set was 0.788 (95% CI: 0.603~0.798), with a sensitivity of 81.3%, specificity of 86.2%, and a calibration curve closely aligned with the ideal curve, yielding a Brier score of 0.115. For the validation set, the AUC was 0.745 (95% CI: 0.668~0.865), with a sensitivity of 83.9% and specificity of 85.5%. SHAP analysis indicated that key predictors of disability status in middle-aged and elderly individuals included age, number of chronic conditions, presence of arthritis, sleep duration, and gender. CONCLUSION: The trajectory prediction model developed in this study demonstrated robust predictive capability for disability-free life expectancy in middle-aged and elderly populations. This model holds potential for supporting early prevention and decision-making in care management.