Abstract
Predicting the likelihood of developing Alzheimer's disease (AD) dementia in at-risk individuals is critical for the design and recruitment of clinical trials for disease-modifying therapies. While machine learning (ML) has shown promise in this area, there is a lack of models specifically for the preclinical AD population—individuals who show early signs of abnormal brain amyloidosis but remain cognitively unimpaired. In this study, we trained and evaluated ML classifiers to predict whether individuals with preclinical AD would progress to mild cognitive impairment (MCI) or dementia within fixed time windows ranging from one to five years. Using regional imaging features from amyloid PET and MRI pooled across seven independent sites and two radiotracers, our models achieved out-of-sample receiver operating characteristic area-under-the-curve (ROC-AUC) values of 0.66 or greater. Furthermore, retroactive cohort enrichment analysis on A4 clinical trial data demonstrated increased statistical power to detect treatment effects. These findings underscore the potential of ML-based patient stratification to improve the efficiency of future clinical trials.
Introduction
Alzheimer’s Disease (AD) clinical trials face a notoriously high failure rate, often because interventions occur too late in the disease progression. Consequently, research has shifted toward the preclinical stage—identifying individuals who are "amyloid-positive" (possessing protein plaques) but still cognitively normal.
However, preclinical AD is highly heterogeneous; while some individuals remain stable for years, others decline rapidly. This variance poses a significant challenge for clinical trials: if a cohort includes too many "stable" individuals, the efficacy of a drug may be masked by a lack of measurable decline in the group. This study addresses this challenge by developing a generalizable machine learning framework to predict progression from preclinical AD to symptomatic stages within one to five years.
Methodology
Machine Learning Framework
The researchers employed Support Vector Machines (SVM) with linear kernels to perform binary classification (Stable vs. Progressor). To capture the nuances of disease progression, they trained five separate models corresponding to different "time-to-progression" windows (1, 2, 3, 4, and 5 years).
Multimodal Feature Integration
The models integrated several data types to maximize predictive accuracy:
- Amyloid PET (SUVR): Regional amyloid signals, rather than just global averages.
- Structural MRI: Regional brain volumes to detect early atrophy.
- Non-Imaging Factors: Age, sex, and the presence of the APOE4 genotype.
Validation and Generalizability
To ensure the models were not overfitted to specific hospital data, the team utilized leave-one-site-out and leave-one-tracer-out cross-validation. This tested the models' ability to generalize across different PET scanners and chemical tracers ([18F]-florbetapir and [11C]-PiB) without specific site-based tuning.
Data Sources
The study pooled data from seven independent sites, creating a robust dataset of 343 stable individuals and 247 progressors. Data sources included:
- A4 Study (Anti-Amyloid Treatment in Asymptomatic Alzheimer’s)
- ADNI (Alzheimer’s Disease Neuroimaging Initiative)
- HABS (Harvard Aging Brain Study)
- MCSA (Mayo Clinic Study of Aging)
- OASIS (Open Access Series of Imaging Studies)
- PAC (Preclinical AD Consortium)
Key Insights
The Value of Regional Patterns
While clinical eligibility often relies on a single "global" amyloid score, this study found that regional patterns—specifically within the temporal and frontal cortices—are significantly more sensitive markers for imminent cognitive decline.
The "Amyloid Lead" Effect
Analysis of feature importance revealed that amyloid PET features became increasingly dominant as the prediction window lengthened (e.g., 5-year vs. 1-year predictions). This suggests that specific amyloid deposition patterns act as an early-warning system that precedes measurable structural brain atrophy captured by MRI.
Clinical Trial Enrichment
A retroactive analysis of the A4 trial provided a proof-of-concept for "cohort enrichment." By using the ML model to filter participants to only those predicted to decline, the researchers demonstrated that treatment effects could be detected with significantly smaller sample sizes. For a sample size of 40 per group, the statistical power to detect differences in amyloid accumulation jumped from ~40% (unenriched) to ~65% (ML-enriched).
Results and Generalizability
The classifiers demonstrated strong performance across diverse datasets:
- Site Generalizability: Out-of-sample ROC-AUC scores ranged from 0.66 to 0.96 for most sites.
- Tracer Agnosticism: When tested on entirely unseen radiotracers, the models maintained an AUC of 0.72 or higher, proving they can function across different clinical environments.
Limitations and Future Directions
Despite the success of the SVM models, the study noted that the preclinical population remains difficult to recruit, leading to smaller sample sizes than those typically used in deep learning. This necessitated the use of traditional ML over complex architectures like Transformers.
Additionally, because each time-window model was trained independently, there is a risk of temporal inconsistency (e.g., predicting stability at 5 years but decline at 2 years). Future research will explore survival SVMs or time-to-event models to ensure longitudinal consistency. Finally, the poor performance at one specific site (HABS) highlighted that extreme class imbalance and small local samples remain a challenge for universal ML application.
Conclusion
As the search for disease-modifying therapies for Alzheimer’s moves toward asymptomatic populations, ML-based stratification offers a path toward more efficient clinical trials. By identifying the individuals most likely to decline, researchers can build more homogeneous cohorts, increasing the statistical power to identify life-changing treatments.
Summary updated