Dual Model Deep Learning for Alzheimer Prognostication

TL;DR

Alzheimer's prognostication at the first visit is hampered by a lack of uncertainty-aware predictions. PROGRESS introduces a dual-model framework that converts baseline CSF biomarker data into calibrated trajectory uncertainty and individualized time-to-conversion predictions without requiring longitudinal history. The trajectory model delivers honest prediction intervals, while the survival model achieves state-of-the-art discrimination (C-index ≈0.83; AUC up to ≈0.88) and demonstrates strong cross-center generalizability after biomarker harmonization. Across a large NACC cohort, PROGRESS yields clinically meaningful risk stratification (seven-fold differences) and shows robustness to center heterogeneity, supporting deployment-oriented, uncertainty-aware personalized prognostication for AD therapies.

Abstract

Disease modifying therapies for Alzheimer's disease demand precise timing decisions, yet current predictive models require longitudinal observations and provide no uncertainty quantification, rendering them impractical at the critical first visit when treatment decisions must be made. We developed PROGRESS (PRognostic Generalization from REsting Static Signatures), a dual-model deep learning framework that transforms a single baseline cerebrospinal fluid biomarker assessment into actionable prognostic estimates without requiring prior clinical history. The framework addresses two complementary clinical questions: a probabilistic trajectory network predicts individualized cognitive decline with calibrated uncertainty bounds achieving near-nominal coverage, enabling honest prognostic communication; and a deep survival model estimates time to conversion from mild cognitive impairment to dementia. Using data from over 3,000 participants across 43 Alzheimer's Disease Research Centers in the National Alzheimer's Coordinating Center database, PROGRESS substantially outperforms Cox proportional hazards, Random Survival Forests, and gradient boosting methods for survival prediction. Risk stratification identifies patient groups with seven-fold differences in conversion rates, enabling clinically meaningful treatment prioritization. Leave-one-center-out validation demonstrates robust generalizability, with survival discrimination remaining strong across held-out sites despite heterogeneous measurement conditions spanning four decades of assay technologies. By combining superior survival prediction with trustworthy trajectory uncertainty quantification, PROGRESS bridges the gap between biomarker measurement and personalized clinical decision-making.

Figures (7)

• Figure 1: NACC Data Integration Pipeline for AD Progression Prediction. The five-phase pipeline systematically processes heterogeneous data sources: (Phase 1) loads CSF biomarker and UDS clinical data; (Phase 2) performs preprocessing including CSF harmonization across assay platforms, clinical data processing, and ATN classification; (Phase 3) conducts temporal alignment between CSF collection dates and clinical visits with a maximum 90-day window; (Phase 4) constructs longitudinal sequences with overlapping windows of length L=5; and (Phase 5) integrates all components into the final dataset. Quality control checkpoints (diamonds) ensure only participants meeting strict criteria proceed to subsequent phases. The resulting integrated dataset comprises 3,051 participants from 43 ADCs with harmonized static biomarkers linked to dynamic clinical trajectories spanning a mean follow-up of 3.9 years.
• Figure 2: Training and validation loss curves for the dual-model PROGRESS framework. Left: Trajectory Parameter Network converging to negative log-likelihood loss, where increasingly negative values indicate improved uncertainty calibration. Right: Deep Survival Network demonstrating monotonic decrease in the combined Cox partial likelihood and pairwise ranking loss. The close tracking between training and validation curves in both models indicates effective regularization preventing overfitting.
• Figure 3: Distribution of performance metrics across hidden layer widths. Box plots show median (orange line), interquartile range (box), and full range (whiskers) across three independent runs. Left: C-index for survival prediction. Center: Intercept R$^2$. Right: Slope R$^2$, showing high variance across all configurations.
• Figure 4: Predicted versus observed trajectory parameters on the held-out test set. Left: Intercept ($\alpha$) predictions ($R^2 = 0.392$, $r = 0.718$). Center: Slope ($\beta$) predictions ($R^2 = 0.056$, $r = 0.240$). Right: Acceleration ($\gamma$) predictions ($R^2 = -0.024$, $r = 0.255$). Red dashed lines indicate perfect prediction.
• Figure 5: Survival model performance visualization. Left: Distribution of predicted risk scores stratified by outcome status, demonstrating separation between censored subjects (blue) and those who experienced conversion events (orange). Right: Scatter plot of risk scores versus observed follow-up time, with color indicating event status.