Development and validation of a parsimonious and noninvasive clinical decision machine learning model for predicting the risk of long-term mild cognitive impairment in older adults: A multinational cohort study.

Abstract

Early identification of mild cognitive impairment (MCI) is crucial for delaying the progression of dementia. However, existing prediction models often require expensive testing or invasive biomarkers, limiting their scalability in primary care settings. Data were drawn from three prospective cohorts: the Chinese Longitudinal Healthy Longevity Survey (CLHLS), the English Longitudinal Study of Ageing (ELSA), and the Health and Retirement Study (HRS). We included 11,069 participants aged ≥60 years who were free of cognitive impairment at baseline. Incident MCI was defined via cohort-specific assessments. The CLHLS was used for internal training and testing, and the ELSA and HRS were used for independent external validation. After feature selection via Shapley additive explanations (SHAP) and recursive feature elimination, 9 machine learning models were developed. Model performance was evaluated on the basis of discrimination, calibration, and clinical utility. SHAP was used for model interpretation. Over the 6-year follow-up period, 2486 participants developed MCI (CLHLS: n = 1008; ELSA: n = 488; HRS: n = 990). Five easily accessible baseline predictors were selected: age, sex, education, instrumental activities of daily living, and baseline cognitive score. The gradient boosting classifier demonstrated favorable overall performance, achieving an internal area under the curve (AUC) of 0.869 (95% CI 0.84-0.90). It maintained robust generalizability during external validation in the ELSA (AUC 0.786, 95% CI 0.76-0.81) and HRS (AUC 0.745, 95% CI 0.73-0.76) cohorts. Furthermore, an interactive web tool incorporating Shapley additive explanations was deployed to conduct a transparent risk assessment. This study developed and externally validated a parsimonious, cross-national machine learning model utilizing five noninvasive features to predict long-term MCI risk. Integrated with a transparent SHAP framework and deployed as an interactive web application, it provides a cost-effective clinical decision support tool for early MCI screening in primary care (https://mcipredictor.streamlit.app/).