Machine learning-based LASSO-Cox model for dementia prediction: The role of midlife cardiometabolic, inflammatory, and genetic risk factors in a US cohort

Abstract

Purpose We aimed to identify key midlife dementia predictors and develop a novel machine learning (ML) -enabled risk prediction model. Methods Using data from 9266 Atherosclerosis Risk in Communities study participants (aged 45–64 years at baseline, 1987–1989). Incident dementia was ascertained through December 2019. A ML-based LASSO-Cox model was applied to develop the risk prediction model. Results Over a 25-year mean follow-up, 2010 participants developed dementia. The LASSO-Cox model identified 12 key predictors and achieved C-indices (95 %CI) of 0.77 (0.75–0.79) in the training set (n = 6182) and 0.78 (0.76–0.81) in the test set (n = 3084). Predictors included age, Digit Symbol Substitution Test, apolipoprotein E ε4, HbA1c, brachial blood pressure, Factor VIII, Delayed Word Recall Test, hypertension, stroke history, C-reactive protein, white blood cell count, and apolipoprotein B. The resulting nomogram demonstrated strong discrimination (AUC 0.77–0.86) and good calibration. LASSO-Cox risk score quartiles effectively stratified participants into low, moderate, high, and very high dementia risk groups. Conclusions The findings demonstrate that the newly developed machine learning-based LASSO-Cox model provides a robust method to predict individuals at high risk of dementia.