Time-to-event prediction using survival analysis methods for Alzheimer's disease progression.

Area of Science:

• Biomedical Informatics
• Computational Neuroscience
• Gerontology

Background:

• Alzheimer's disease (AD) detection and progression are well-researched, but continuous-time survival prediction remains underexplored.
• Predictive analytics for AD progression can support medical practitioners in patient management.
• Current methods lack robust tools for forecasting the next stage of AD over time.

Purpose of the Study:

• To develop a survival analysis approach for predicting the probability of AD next stage progression.
• To examine interactions between temporal and medical patterns for personalized AD forecasting.
• To provide medical practitioners with tools for timely analysis and personalized treatment recommendations.

Main Methods:

• Simulated disease progression using non-linear survival models: non-linear Cox proportional hazard model (Cox-PH) and neural multi-task logistic regression (N-MTLR).
• Evaluated model performance using concordance index (C-index) and Integrated Brier Score (IBS).
• Developed deep neural network models using National Alzheimer's Coordinating Center data (2005-2017) with multiple visit details.

Main Results:

• N-MTLR based survival models outperformed CoxPH models, achieving a C-index of 0.79 and IBS of 0.09.
• Identified 50 critical features from 92 using recursive feature elimination and random forest, including cognition, behavior, and dementia criteria.
• Feature selection demonstrated improved probability prediction effectiveness at each time interval.

Conclusions:

• The deep learning-based survival method enables efficient prediction of AD stage shifts for personalized treatment.
• The approach can help mitigate or postpone the effects of Alzheimer's disease.
• The survival analysis framework is adaptable for predicting disease stage shifts in other progressive conditions like cancer and Huntington's disease.