[Paper Review] Prediction of cognitive decline for enrichment of Alzheimer's disease clinical trials

Angela Tam, Alexis Laurent+2 more|arXiv (Cornell University)|Nov 7, 2021

Dementia and Cognitive Impairment Research24 citations

TL;DR

This study develops machine learning models that predict cognitive decline in Alzheimer’s disease patients and presymptomatic individuals using baseline data—demographics, cognitive tests, APOE status, and MRI—achieving up to 79% AUC. The models enable trial enrichment by selecting high-risk individuals, reducing required sample sizes by up to 51% while maintaining statistical power.

ABSTRACT

A key issue to Alzheimer's disease clinical trial failures is poor participant selection. Participants have heterogeneous cognitive trajectories and many do not decline during trials, which reduces a study's power to detect treatment effects. Trials need enrichment strategies to enroll individuals who will decline. We developed machine learning models to predict cognitive trajectories in participants with early Alzheimer's disease (n=1342) and presymptomatic individuals (n=756) over 24 and 48 months respectively. Baseline magnetic resonance imaging, cognitive tests, demographics, and APOE genotype were used to classify decliners, measured by an increase in CDR-Sum of Boxes, and non-decliners with up to 79% area under the curve (cross-validated and out-of-sample). Using these prognostic models to recruit enriched cohorts of decliners can reduce required sample sizes by as much as 51%, while maintaining the same detection power, and thus may improve trial quality, derisk endpoint failures, and accelerate therapeutic development in Alzheimer's disease.

Motivation & Objective

• Address the high failure rate in Alzheimer’s disease clinical trials due to poor participant selection and heterogeneous cognitive trajectories.
• Improve trial power by identifying individuals most likely to decline during trial periods using predictive modeling.
• Develop prognostic models applicable across the disease spectrum—early Alzheimer’s disease (MCI/dementia) and presymptomatic stages.
• Validate models on independent datasets to ensure generalizability and robustness.
• Demonstrate that model-based enrichment can significantly reduce required sample sizes without sacrificing statistical power.

Proposed method

• Train two separate prognostic machine learning models: one for early Alzheimer’s (24-month decline prediction) and one for presymptomatic individuals (48-month decline prediction).
• Use baseline data including demographics, cognitive test scores (e.g., CDR-SB), APOE ε4 status, and structural MRI scans as input features.
• Train models on combined data from ADNI and NACC, with the early AD model trained on 1,151 individuals and the presymptomatic model on 628 individuals.
• Validate the early AD model on a placebo arm of a Phase 3 clinical trial (n=115) and the PharmaCog dataset (n=76); validate the presymptomatic model on OASIS-3 (n=128).
• Use cross-validation and out-of-sample testing to estimate model performance, reporting area under the ROC curve (AUC).
• Conduct power analyses comparing unenriched vs. enriched cohorts to quantify sample size reduction potential.

Experimental results

Research questions

• RQ1 Can machine learning models trained on routine clinical data predict cognitive decline in early Alzheimer’s disease patients over 24 months with high accuracy?
• RQ2 Can similar models predict decline in cognitively unimpaired individuals over a 48-month window using accessible biomarkers?
• RQ3 How well do these models generalize to independent, real-world clinical trial and imaging datasets?
• RQ4 To what extent can model-based enrichment reduce required sample sizes in Alzheimer’s disease clinical trials while preserving statistical power?
• RQ5 Can prognostic models improve trial quality by reducing endpoint failure risk due to poor participant selection?

Key findings

• The prognostic model for early Alzheimer’s disease achieved an area under the curve (AUC) of up to 79% in cross-validated and out-of-sample evaluations.
• The model for presymptomatic individuals also demonstrated strong predictive performance, with validation results reported on the OASIS-3 cohort.
• Power analyses showed that using the model to enrich clinical trial cohorts with predicted decliners could reduce required sample sizes by up to 51% while maintaining the same statistical power.
• The early AD model generalized well to an independent placebo arm of a Phase 3 clinical trial, indicating robustness to real-world trial conditions.
• The models outperform traditional inclusion criteria by integrating multiple biomarkers and clinical features into a holistic risk prediction framework.
• The approach enables precision recruitment by identifying high-risk individuals earlier and more accurately than APOE ε4 status alone.