Predicting future cognitive impairment in preclinical Alzheimer's disease using amyloid PET and MRI: a multisite machine learning study

This multisite machine learning study demonstrates that models trained on amyloid PET and MRI data can effectively predict future cognitive decline in preclinical Alzheimer's disease with strong generalizability across sites and tracers, thereby enhancing statistical power for clinical trial recruitment and treatment effect detection.

The Gist

The Big Picture: Predicting the "Slow Leak" Before the "Flood"

Imagine Alzheimer's disease not as a sudden flood that destroys a house, but as a slow, invisible leak in the roof. For years, water (amyloid plaques) drips into the attic, damaging the structure, but the people living downstairs (the brain's cognitive functions) don't notice anything is wrong yet. They are still cooking, talking, and remembering names perfectly. This is the preclinical stage.

The problem is that by the time the water starts dripping onto the living room floor (memory loss and dementia), it's often too late to fix the roof easily. Scientists want to catch these leaks early. But here's the tricky part: not every leak turns into a flood. Some houses have a slow drip that never causes damage, while others are about to collapse.

This paper is about building a super-smart "Leak Detector" (Machine Learning) that can look at a house's blueprint and tell you which roofs are about to fail in the next 1, 2, 3, 4, or 5 years.

---

How They Built the Detector

The researchers didn't just look at one part of the house. They used two powerful tools to scan the brain:

1. The "Smoke Detector" (Amyloid PET Scan): This lights up where the "amyloid" (the sticky gunk causing the leak) is piling up.
2. The "Structural Scan" (MRI): This measures the size of the rooms. In Alzheimer's, the brain shrinks (atrophies), like a house losing its volume as the walls crumble.

They fed data from 7 different research centers (like 7 different neighborhoods) and used two different types of scanners into a computer program called a Support Vector Machine (SVM). Think of the SVM as a very strict, super-observant teacher who looks at thousands of student test scores (brain scans) to learn the pattern of who is going to fail the class (develop dementia) and who will pass.

The "Test Drive": Does It Work on New Students?

Usually, when you train a robot to recognize cats, you show it a million pictures of cats. But if you only show it fluffy white cats, it might get confused when it sees a black cat.

To make sure their "Leak Detector" was truly smart and not just memorizing the specific houses it studied, they did a strict test:

• They trained the model on data from 6 sites.
• They tested it on the 7th site it had never seen before.
• They did the same thing with the different scanner types.

The Result: The model was surprisingly good! It could predict who would develop cognitive problems with decent accuracy, even on data it had never seen. It was like a weather forecaster who could predict rain in a city they had never visited, just by looking at the clouds.

The "A4" Experiment: Finding the Right People for a Cure

The ultimate goal of this isn't just to predict the future; it's to help cure the disease.

Imagine a pharmaceutical company is testing a new medicine to stop the roof leak. They need to test it on people who are actually going to get sick soon. If they test it on people whose roofs are perfectly fine and will never leak, the medicine will look like it does nothing (because there was no leak to stop).

The researchers took data from a real clinical trial called the A4 study (which tested a drug called Solanezumab). They ran their "Leak Detector" on the participants after the trial was over.

• Before: The trial had a mix of people. Some were about to crash, some were fine. The results were muddy.
• After (Enrichment): They used the AI to pick out only the people the AI predicted were "high risk" (the ones with the leak about to burst).

The Finding: When they looked at just this "high-risk" group, the signal of the drug working became much clearer (specifically regarding the amount of amyloid in the brain). It's like trying to hear a whisper in a noisy room. If you ask everyone to be quiet, you can't hear the whisper. But if you ask only the people who are whispering to speak up, you can hear them perfectly.

The "Secret Ingredients": What Did the AI Learn?

The researchers asked the AI: "What clues did you use to make your prediction?"

• The Amyloid Map: The AI didn't just care about how much gunk was in the brain; it cared about where it was. It noticed that gunk in the temporal and frontal lobes (the brain's memory and planning centers) was a huge red flag.
• The Shrinking Rooms: It also looked at how much the brain had shrunk, particularly in the ventricles (fluid-filled spaces that get bigger as the brain shrinks).
• The Time Factor: Interestingly, the further out they tried to predict (e.g., 5 years vs. 1 year), the more important the "Amyloid Map" became. It's like saying, "If you want to know if a house will collapse in 5 years, look at the foundation cracks (amyloid). If you want to know if it will collapse tomorrow, look at the crumbling walls (brain shrinkage)."

The Bottom Line

This paper is a hopeful step forward. It shows that we can use AI to look at brain scans and say, "Hey, this person looks healthy right now, but their brain is showing early warning signs that they might struggle in the next few years."

This is a game-changer for clinical trials. Instead of guessing who to test a new drug on, doctors can use this AI to find the "perfect candidates"—people who are healthy now but are on the fast track to decline. This makes it much easier to see if a new medicine actually works, potentially speeding up the discovery of cures for Alzheimer's.

In short: They built a crystal ball made of math and MRI machines that helps us find the right people to save before the damage is done.

Technical Summary

1. Problem Statement

The development of disease-modifying therapies for Alzheimer's Disease (AD) has historically faced high failure rates, often attributed to recruiting participants at too advanced a disease stage or enrolling heterogeneous cohorts with varying rates of cognitive decline. Recent clinical trials have shifted focus to preclinical AD (asymptomatic individuals with abnormal brain amyloidosis) to intervene earlier. However, predicting which preclinical individuals will progress to Mild Cognitive Impairment (MCI) or dementia remains challenging due to:

• Heterogeneity: Variable rates of decline among amyloid-positive, cognitively unimpaired (CU) individuals.
• Data Variability: Differences in imaging protocols, scanner models across multiple sites, and the use of different amyloid PET radiotracers (e.g., [18F]-florbetapir vs. [11C]-Pittsburgh-compound-B).
• Lack of Generalizable Models: Existing Machine Learning (ML) models often focus on symptomatic populations, rely on single time-windows, or fail to generalize across external sites and tracers.

2. Methodology

Data Sources and Cohort

• Datasets: The study pooled data from seven independent sites/consortia: A4 study, ADNI, HABS, MCSA, OASIS, and the Preclinical Alzheimer's Disease Consortium (PAC).
• Participants: 590 total participants (343 stable, 247 progressors).
  • Inclusion: Cognitively unimpaired (Global CDR = 0) at baseline with amyloid-positive PET scans.
  • Labels:
    • Stable: Remained CDR = 0 for at least 5 years.
    • Progressor: Converted to CDR > 0 within 1 to 5 years.
• Features:
  • Imaging: Regional Amyloid PET Standardized Uptake Value Ratios (SUVR) from 114 cortical/subcortical regions and MRI volumetric features from 125 regions.
  • Non-Imaging: Age, sex, and APOE4 carrier status.

Model Development

• Algorithm: Support Vector Machines (SVM) with a linear kernel.
• Task: Binary classification (Stable vs. Progressor) for five independent time-to-progression windows (1, 2, 3, 4, and 5 years).
• Training Strategy:
  • Class Balancing: Employed class-balanced weighting to penalize misclassification of the minority class (progressors) more heavily.
  • Feature Preprocessing: Z-scored using the training fold's stable participants only.
  • Hyperparameter Tuning: Grid search with 5-fold cross-validation to optimize the regularization parameter $C$.

Evaluation Framework

• Cross-Validation: Rigorous Leave-One-Site-Out (LOSO) and Leave-One-Tracer-Out (LOTO) cross-validation.
  • LOSO: Trained on 6 sites, tested on the 7th.
  • LOTO: Trained on one tracer (e.g., PiB), tested on the other (e.g., FBP).
• Sensitivity Analysis: Retrained models using Centiloid-harmonized SUVRs to assess the impact of tracer standardization.
• Feature Importance:
  • Modality Importance: Nested models (omitting PET, MRI, or non-imaging features) to measure ROC-AUC drop.
  • Feature-Level Importance: Covariance-corrected linear SVM weights (Haufe transformation) to interpret feature contributions while accounting for collinearity.

Clinical Trial Application (A4 Study)

• Validation: Applied the model trained on all external sites to the A4 trial cohort (Placebo and Solanezumab arms).
• Cohort Enrichment: Retroactively stratified the A4 cohort into "predicted progressors" and "predicted stables."
• Outcome Analysis:
  • Compared longitudinal Preclinical Alzheimer's Cognitive Composite (PACC) scores between predicted groups.
  • Tested for treatment effects (Placebo vs. Solanezumab) on PACC and Amyloid PET SUVR changes in the enriched vs. unenriched cohorts.
  • Power Analysis: Bootstrap analysis (1,000 iterations) to estimate statistical power for detecting treatment effects at various sample sizes.

3. Key Results

Model Performance and Generalizability

• Site Generalizability: Models achieved an out-of-sample ROC-AUC $\ge$ 0.66 for all hold-out sites except HABS (which had low sample size and variability, resulting in AUC < 0.5). Most sites exceeded 0.75.
• Tracer Generalizability: Models trained on one tracer achieved ROC-AUC $\ge$ 0.72 when tested on the other tracer.
  • Observation: Without harmonization, sensitivity/specificity trade-offs were skewed (e.g., high specificity but low sensitivity when testing FBP on a PiB-trained model).
  • Harmonization: Converting to Centiloids improved the balance between sensitivity and specificity, though ROC-AUC changes were not statistically significant.
• Metrics: High specificity (>0.74) and Negative Predictive Value (>0.74) were consistent across time windows. Sensitivity and PPV were lower for shorter time windows (1-2 years) due to class imbalance.

Feature Importance

• Amyloid PET: The most critical modality, especially for longer time horizons (3-5 years). Key regions included bilateral inferior/middle temporal gyri, middle/superior frontal gyri, and middle occipital gyri.
• MRI: Ventricular volumes (especially inferior lateral ventricles) and hippocampal/amygdala volumes were highly important.
• Non-Imaging: Age was the strongest non-imaging predictor; APOE4 and sex showed weak associations.
• Temporal Dynamics: The importance of amyloid PET features increased significantly as the time-to-progression window expanded, suggesting amyloid distribution is a stronger predictor of early decline than volumetric atrophy in the preclinical stage.

A4 Trial Enrichment Analysis

• Stratification Validity: Predicted progressors showed significantly faster cognitive decline (lower PACC scores) compared to predicted stables in both placebo and treatment arms.
• Treatment Effect Detection:
  • Primary Outcome (PACC): No significant treatment effect was found in either the unenriched or enriched cohorts (consistent with the original A4 trial results, suggesting Solanezumab may lack efficacy).
  • Secondary Outcome (Amyloid SUVR): Enrichment using 2, 4, and 5-year models increased statistical power to detect differences in amyloid accumulation between Placebo and Solanezumab arms.
  • Power Gain: For sample sizes between 20–80, the enriched cohorts showed higher power to detect the treatment effect on amyloid reduction compared to the unenriched cohort. The 1-year and 3-year models did not show this benefit (likely due to sample size constraints or cohort heterogeneity).

4. Key Contributions

1. Multisite Generalizability: Demonstrated that ML models trained on amyloid PET and MRI can generalize across diverse clinical sites and different PET radiotracers without explicit harmonization (though Centiloids improve metric balance).
2. Preclinical Specificity: Focused exclusively on the amyloid-positive, cognitively unimpaired population, addressing a critical gap in the literature which often focuses on MCI/AD.
3. Clinical Utility: Provided empirical evidence that ML-based stratification can enhance the statistical power of clinical trials by enriching cohorts with individuals most likely to progress, specifically for secondary endpoints like amyloid accumulation.
4. Feature Insights: Identified that regional amyloid patterns (specifically in temporal and frontal cortices) are more predictive of future decline in preclinical stages than global amyloid burden or subcortical atrophy alone.

5. Significance and Future Directions

• Clinical Trials: This study supports the use of multimodal ML models for cohort enrichment in future AD prevention trials. By selecting "fast decliners," trials can reduce sample size requirements and increase the likelihood of detecting disease-modifying effects.
• Prognostic Tools: The models serve as potential prognostic tools for identifying high-risk individuals for early intervention, provided amyloid PET becomes more accessible for asymptomatic screening.
• Limitations & Future Work:
  • Tau PET: The study did not include tau PET, which is a strong predictor of decline. Future models should integrate tau imaging.
  • Sample Size: Class imbalance (few progressors) limits deep learning applications; larger cohorts are needed.
  • Temporal Consistency: Current models treat time windows independently. Future work should explore time-to-event models (e.g., Survival SVM) to ensure predictions are temporally consistent.

In conclusion, the authors successfully developed and validated robust machine learning classifiers that can predict cognitive decline in preclinical AD across diverse datasets, offering a viable strategy to optimize the design and statistical power of future Alzheimer's clinical trials.