Joint ensemble learning-based risk prediction of Alzheimer's disease among mild cognitive impairment patients

Due to the recognition for the importance of early intervention in Alzheimer's disease (AD), it is important to focus on prevention and treatment strategies for mild cognitive impairment (MCI). This study aimed to establish a risk prediction model for AD among MCI patients to provide clinical guidance for primary medical institutions.

Data from MCI subjects were obtained from the NACC. Importance ranking and the SHapley Additive exPlanations (SHAP) method for the Random Survival Forest (RSF) and Extreme Gradient Boosting (XGBoost) algorithms in ensemble learning were adopted to select the predictors, and hierarchical clustering analysis was used to mitigate multicollinearity. The RSF, XGBoost and Cox proportional hazard regression (Cox) models were established to predict the risk of AD among MCI patients. Additionally, the effects of the three models were evaluated.

A total of 3674 subjects with MCI were included. Thirteen predictors were ultimately identified. In the validation set, the concordance indices were 0.781 (RSF), 0.781 (XGBoost), and 0.798 (Cox), and the Integrated Brier Score was 0.087 (Cox). The prediction effects of the XGBoost and RSF models were not better than those of the Cox model.

The ensemble learning method can effectively select predictors of AD risk among MCI subjects. The Cox proportional hazards regression model could be used in primary medical institutions to rapidly screen for the risk of AD among MCI patients once the model is fully clinically validated. The predictors were easy to explain and obtain, and the prediction of AD was accurate.