Dual-model deep learning for Alzheimer's prognostication.

Area of Science:

• Neuroscience
• Biomarker Discovery
• Artificial Intelligence in Medicine

Background:

• Current Alzheimer's disease (AD) predictive models lack practicality for initial patient encounters due to requirements for longitudinal data and absence of uncertainty quantification.
• Accurate prognostic timelines are crucial for timely administration of disease-modifying therapies in AD.
• The need for robust, generalizable prognostic tools that utilize single-time-point biomarker data is critical for clinical decision-making.

Purpose of the Study:

• To develop and validate a novel deep learning framework, PROGRESS (PRognostic Generalization from REsting Static Signatures), for predicting Alzheimer's disease progression from baseline cerebrospinal fluid (CSF) biomarkers.
• To provide accurate, uncertainty-quantified prognostic estimates for individualized cognitive decline and time-to-dementia conversion.
• To overcome the limitations of existing predictive models by enabling prognostic assessment at the first clinical visit without prior patient history.

Main Methods:

• Developed a dual-model deep learning framework (PROGRESS) integrating a probabilistic trajectory network and a deep survival model.
• Utilized a large dataset of over 3000 participants from the National Alzheimer's Coordinating Center (NACC) database across 43 Alzheimer's Disease Research Centers.
• Employed leave-one-center-out cross-validation to assess generalizability across heterogeneous measurement conditions and assay technologies spanning decades.

Main Results:

• PROGRESS demonstrated superior survival prediction performance compared to traditional methods like Cox proportional hazards, Random Survival Forests, and gradient boosting.
• The probabilistic trajectory network provided calibrated uncertainty bounds for individualized cognitive decline parameters, achieving near-nominal coverage.
• Risk stratification identified patient subgroups with substantial differences in conversion rates (up to seven-fold), enabling effective treatment prioritization. Survival discrimination remained robust across diverse clinical sites.

Conclusions:

• PROGRESS offers a significant advancement in Alzheimer's disease prognostication by transforming single CSF biomarker assessments into actionable, uncertainty-quantified estimates.
• The framework enables honest prognostic communication and supports personalized clinical decision-making at the critical first-visit encounter.
• PROGRESS bridges the gap between biomarker data and clinical utility, providing essential prognostic timelines currently lacking in Alzheimer's disease staging.