Supplementary Material for: Prediction of Cognitive Impairment Risk among Older Adults: A Machine-Learning Based Comparative Study and Model Development

Detecting cognitive decline early in the older adults is crucial for effective intervention. This study, part of the Ma'anshan Healthy Aging Cohort Study, examined 2,288 participants with normal cognitive function. Forty-two potential predictors, including demographics, chronic diseases, lifestyle factors, and baseline cognitive function, were selected. The dataset was divided into training, validation, and test sets (60%, 20%, and 20%, respectively). Recursive feature elimination (RFE) and six machine learning algorithms were used for model development. Model performance was assessed using area under the curve (AUC), specificity, sensitivity, and accuracy. SHapley Additive exPlanations (SHAP) was applied for interpretability, revealing the top ten influential features: baseline MMSE, education, economic status, social activities, PSQI, BMI, SBP, DBP, IADL, and age. The Naïve Bayes (NB) algorithm-based model achieved an AUC of 0.820 (95% CI 0.773-0.887) on the test set, outperforming other algorithms. This model can help primary healthcare staff in community settings identify individuals at higher risk of cognitive impairment within three years among older adults.

及早识别老年人群的认知衰退，对于开展有效干预至关重要。本研究作为马鞍山健康老年队列研究（Ma'anshan Healthy Aging Cohort Study）的一部分，共纳入2288名认知功能正常的受试者。研究筛选出42项潜在预测因子，涵盖人口统计学特征、慢性病状况、生活方式因素以及基线认知功能水平。本数据集按60%、20%、20%的比例划分为训练集、验证集与测试集。采用递归特征消除法（Recursive Feature Elimination, RFE）与6种机器学习算法开展模型构建。模型性能通过曲线下面积（Area Under the Curve, AUC）、特异度、敏感度与准确率进行评估。采用SHapley可加解释法（SHapley Additive exPlanations, SHAP）进行模型可解释性分析，筛选出排名前十的影响特征：基线简易精神状态检查表（Mini-Mental State Examination, MMSE）得分、受教育程度、经济状况、社交活动、匹兹堡睡眠质量指数（Pittsburgh Sleep Quality Index, PSQI）、体重指数（Body Mass Index, BMI）、收缩压（Systolic Blood Pressure, SBP）、舒张压（Diastolic Blood Pressure, DBP）、工具性日常生活活动能力（Instrumental Activities of Daily Living, IADL）与年龄。基于朴素贝叶斯（Naïve Bayes, NB）算法的模型在测试集上取得了0.820的曲线下面积（95%置信区间：0.773~0.887），性能优于其余算法。该模型可帮助社区基层医疗人员识别老年人群中未来三年内认知障碍高风险个体。