Neural Network-Based Dynamic Prediction for Interval-Censored Data With Time-Varying Covariates: Application to Alzheimer's Disease.

Alzheimer's disease (AD) is a progressive neurodegenerative disorder accounting for a significant proportion of global dementia cases. Given the lack of effective treatments, there is growing interest in dynamic prediction methods for timely interventions. Notably, many at-risk individuals with periodic clinic visits provide dynamic cognitive and functional scores. When an individual receives a new score at each follow-up, the dynamic prediction model can integrate the individual's historical scores with the new follow-up scores to offer an updated risk prediction. This study utilizes a comprehensive dataset from the four phases of the Alzheimer's Disease Neuroimaging Initiative (ADNI) study, comprising 1702 individuals with multiple time-varying cognitive and functional scores and baseline covariates. We address several challenges: Interval-censored time-to-AD due to intermittent assessments, multiple time-varying covariates, and nonlinear covariate effects on AD development. The proposed approach integrates multivariate functional principal component analysis with a neural network; the former extracts important predictive features from multiple time-varying covariates, while the latter handles the nonlinear covariate effects on interval-censored time-to-AD. This method facilitates individualized and dynamic predictions for AD development. Based on simulation results and application to the ADNI dataset, the proposed method outperforms several other methods in terms of prediction accuracy. Furthermore, it identifies high- and low-risk subgroups with distinct progression risk profiles at each landmark time, enabling early and timely intervention of AD. To facilitate dynamic predictions in practice, we have developed an online prediction platform accessible at http://olap.ruc.edu.cn.