Predicting mortality risk in Alzheimer's disease using machine learning based on lifestyle and physical activity.

Alzheimer's disease (AD), a progressive neurodegenerative disorder, significantly impacts patient survival, prompting the need for accurate prognostic tools. Lifestyle factors and physical activity levels have been identified as critical modifiable risk factors influencing AD outcomes, but their precise impact on mortality prediction remains understudied. This study aimed to employ machine learning (ML) techniques to predict mortality risk in AD patients, leveraging data on lifestyle and physical activity to enhance personalized care strategies and inform public health policies. We analyzed data from 53,231 participants collected from the National Health and Nutrition Examination Survey (NHANES, 2007-2020). Participants were stratified by AD symptom severity using Patient Health Questionnaire-9 scores. Random Survival Forest (RSF) and Cox proportional hazards models were developed and validated using a training set (n = 42,585) and test set (n = 10,646). Model performance was evaluated using the integrated area under the curve (iAUC), integrated Brier score/prediction error (iBS/PE), and concordance index (C-index). The RSF model outperformed the Cox model, achieving higher discrimination and calibration. Specifically, the RSF demonstrated an iAUC of 0.781 (95% CI 0.778-0.839), iBS/PE of 0.150 (95% CI 0.083-0.122), and a C-index of 0.785 (95% CI 0.776-0.800) in the no-symptom group of the training cohort. These metrics indicate superior predictive accuracy, especially at extreme ends of risk prediction. Lifestyle and physical activity levels were identified as significant predictors influencing mortality risk. ML algorithms, notably RSF, effectively predict mortality risk in AD patients, demonstrating clear advantages over traditional statistical models. Incorporating lifestyle and physical activity into ML-based predictive frameworks can significantly improve risk stratification, informing targeted interventions. Further external validation across diverse populations is necessary to establish broader applicability.