Establishment of a mild cognitive impairment risk model in middle-aged and older adults: a longitudinal study.

Abstract

Early identification individuals at high risk of mild cognitive impairment (MCI) is essential for prevention and intervention strategies of dementia, such as Alzheimer's disease. MCI prediction considering the interdependence of predictors in longitudinal data needs to be further explored. We aimed to employ machine learning (ML) to develop and verify a prediction model of MCI. In a longitudinal population-based cohort of China Health and Retirement Longitudinal Study (CHARLS), 8390 non-MCI participants were enrolled. The diagnosis of MCI was based on the aging-associated cognitive decline (AACD), and 13 factors (gender, education, marital status, residence, diabetes, hypertension, depression, hearing impairment, social isolation, physical activity, drinking status, body mass index and expenditure) were finally selected as predictors. We implemented a long short-term memory (LSTM) to predict the MCI risks in middle-aged and older adults within 7years. The Receiver Operating Characteristic curve (ROC) and calibration curve were used to evaluate the performance of the model. Through 7years of follow-up, 1925 participants developed MCI. The model for all incident MCI achieved an AUC of 0.774, and its deployment to the participants followed 2, 4, and 7years achieved results of 0.739, 0.747, and 0.750, respectively. The model was well-calibrated with predicted probabilities plotted against the observed proportions of cognitive impairment. Education level, gender, marital status, and depression contributed most to the prediction of MCI. This model could be widely applied to medical institutions, even in the community, to identify middle-aged and older adults at high risk of MCI.