AI-driven fusion of neurological work-up for assessment of biological Alzheimer’s disease

Varuna Jasodanand, Sahana S Kowshik, Shreyas Puducheri, Michael F. Romano, Lingyi Xu, Rhoda Au, Vijaya B. KolachalamamedRxiv 2025 · DOI

TL;DR

Researchers have developed a transformer-based machine learning framework that predicts the presence of Alzheimer’s biomarkers—amyloid-beta and tau—using standard clinical data like MRIs and cognitive tests. By accurately "estimating" PET scan results without the actual $3,000 procedure, this model could save millions in clinical trial costs and broaden access to early diagnosis.

What Problem Does This Paper Address?

The "gold standard" for diagnosing Alzheimer’s disease (AD) involves Positron Emission Tomography (PET) scans to detect amyloid-beta (Aβ) plaques and tau (τ) tangles in the brain. However, PET imaging is a massive bottleneck: it is expensive, requires specialized facilities, and tau PET remains largely restricted to research settings.

This creates a "screening crisis" in drug development. Between 1995 and 2021, AD research saw a staggering 95% failure rate and $42.5 billion in expenditures, much of which was spent simply screening participants to see if they had the biological markers required for a trial. There is an urgent need for a "rule-out" tool—a way to use routine, accessible data (like medical history and structural MRIs) to identify who actually needs an expensive PET scan and who doesn't.

Methods

The authors propose a transformer-based multimodal fusion framework. While transformers are famous for powering Large Language Models (LLMs), here they are used to integrate disparate data types:

  • Data Modalities: Demographics, medical history, neuropsychological scores (like the MMSE), APOE-ε4 genetics, and structural MRI features.
  • Feature Masking: A key innovation is the use of "random feature masking" during training. This allows the model to remain robust even when a patient is missing certain data points (e.g., if a specific cognitive test wasn't performed), a common occurrence in real-world clinical settings.
  • MRI Embeddings: Instead of just using simple volume measurements, the model uses a "Swin UNETR" architecture to extract deep spatial features from 3D MRI scans.
  • Interpretability: The team used Shapley values (a concept from game theory) to determine which features most influenced the model's decisions and built graph networks to ensure the model’s "attention" aligned with known biological patterns of disease progression.

Datasets

The study is impressively large-scale, leveraging data from seven distinct cohorts totaling 12,185 participants:

  • Training/Internal Validation: NACC (National Alzheimer’s Coordinating Center), A4, OASIS-3, AIBL, and the Framingham Heart Study.
  • External Testing: ADNI (Alzheimer’s Disease Neuroimaging Initiative) and HABS (Harvard Aging Brain Study).
  • Neuropathological Validation: A subset of participants with post-mortem brain examinations to confirm the model's accuracy against physical tissue samples.

Key Insights and Ideas

  • The "Rule-Out" Power: The model achieved a high Negative Predictive Value (NPV)—up to 91.65% for tau. This means if the model says you are "negative," it is very likely correct. In a clinical trial setting, this allows researchers to safely skip PET scans for low-risk individuals.
  • MRI is the Heavy Lifter for Tau: While history and demographics provided a baseline, adding MRI data caused a massive jump in tau prediction accuracy (AUROC from 0.53 to 0.74). This suggests that structural changes in the brain are strong proxies for tau protein accumulation.
  • Biological Alignment: The model wasn't just "guessing" based on correlations. The brain regions the model flagged as "important" (via Shapley values) closely matched the Braak stages—the stereotypical spatial path tau takes as it spreads through the brain.
  • Cost Efficiency: The authors estimate that using this model as a pre-screening tool in their test cohort could have saved over $3.5 million by avoiding unnecessary PET scans.

Key Results

  • Classification Accuracy: The model achieved an AUROC of 0.79 for Aβ and 0.84 for tau status.
  • Correlation with Severity: Model-predicted probabilities showed a strong positive correlation with actual "Centiloid" values (a universal scale for amyloid) and regional tau SUVr (uptake ratios).
  • Post-Mortem Success: The model’s predictions aligned significantly with actual pathological scores (Thal phases, Braak stages, and CERAD scores) observed in autopsied brains, proving it tracks the underlying biology, not just imaging artifacts.

Limitations and Future Work

  • Demographic Bias: The current datasets were predominantly composed of White participants. The authors acknowledge that the model’s generalizability across diverse racial and ethnic populations remains to be fully proven.
  • Binary vs. Continuous: The model currently focuses on binary classification (Positive vs. Negative). Future work aims to move toward ordinal regression, which could predict the specific stage or "load" of the disease more precisely.
  • Blood-Based Biomarkers: While the model can incorporate plasma data (like p-tau 217), the sample size for these features was small in this study. Integrating emerging high-sensitivity blood tests will be a primary focus for future iterations.

Summary updated

Abstract

Abstract Alzheimer’s disease (AD) diagnosis hinges on detecting amyloid beta (A β ) plaques and neurofibrillary tau ( τ ) tangles. While amyloid PET imaging is now clinically approved, tau PET remains largely restricted to research settings. These imaging techniques, though valuable, are expensive and often difficult to access, limiting their widespread use in routine clinical practice. Here, we introduce a computational framework that leverages multimodal data from seven distinct cohorts comprising 12, 185 participants to estimate indi-vidual PET profiles, both global and regional, using more accessible data modalities, such as demographics, medical history, medication use, fluid measurements, functional and neuropsychological assessments, and structural MRIs. Our approach achieved an area under the receiver operating characteristic curve of 0.79 and 0.84 in classifying persons with positive A β and τ status, respectively. Model predictions were consis-tent with various biomarker and cognitive profiles, as well as with different degrees of protein abnormalities observed in post-mortem examinations. Furthermore, the regional volumes identified by the model as im-portant aligned with the spatial distributions of the standardized uptake value ratio for regional τ labels. Our model offers a practical approach to identify potential candidates for newly approved anti-amyloid treatments and AD clinical trials for combined amyloid and tau therapies by utilizing standard neurological evaluation data.
BibTeX
@misc{jasodanand2025ai,
  author = {Varuna Jasodanand and Sahana S Kowshik and Shreyas Puducheri and Michael F. Romano and Lingyi Xu and Rhoda Au and Vijaya B. Kolachalama},
  title = {{AI-driven fusion of neurological work-up for assessment of biological Alzheimer’s disease}},
  year = {2025},
  doi = {10.1101/2025.03.12.25323862},
  howpublished = {medRxiv},
}
Figure