Application of the MoCA score, sociodemographic and health-related factors to evaluate interpretable screening models for geriatric cognitive disorders in Chongqing: a cross-sectional survey study

Geriatric cognitive disorders (GCDs) exhibit a significant age dependence. As global populations age, China, with its substantial demographic, must prioritize the increasing incidence of cognitive disorders among older adults. Notably, the prevalence rate of cognitive disorders in Chongqing is the highest in the country. Social and health-related factors have drawn considerable attention as predictors of cognitive decline because they are noninvasive and readily available. In this context, this study uses MoCA score, sociodemographic and health-related data to develop and validate machine-learning models that predict cognitive disorders in the elderly in Chongqing city and assess individual classification. We conducted a cross-sectional survey of older adults in Chongqing using stratified cluster sampling to collect sociodemographic data (4 items), health-related variables (13 items) and MoCA score. We then trained nine machine-learning algorithms to build screening models for GCD. The aims were to clarify the predictors associated with prevalent GCD and to develop early screening models for GCD. We evaluated model performance by area under the curve (AUC), calibration curve, decision curve analysis, sensitivity, specificity, and validation cohorts. The Shapley additive interpretation (SHAP) method is used to rank the importance of features and interpret the final model. The prevalence of cognitive disorders among the elderly in Chongqing (32.8%) significantly surpasses the national average. Through feature selection, 11 variables were identified as independent predictors of GCD. Prediction models were developed based on these factors through nine distinct machine learning methods. Among the nine ML models, the XGB model achieved a relatively high AUC value (0.9397, 95%CI 0.9136–0.9658). External validation further confirmed the validity of the XGB model, with an AUC value of 0.9349 (95%CI 0.9139–0.9558). According to SHAP analysis, age, high education, MoCA score, hyperglycemia, hypertension, marital status, cerebral infarction, and cerebral atrophy are the key predictors for screening the likelihood of cognitive disorders in the elderly. We developed a machine learning–based prediction model that indicates potential screening utility to assess the risk of cognitive disorders in older adults.