Predicting ARIA incidence due to anti-amyloid therapy for Alzheimer’s disease with multimodal risk modeling frameworks
Abstract Amyloid-related imaging abnormalities (ARIA) are potentially serious adverse side effects of amyloid-targeting treatment (ATT) for Alzheimer’s disease. Pre-screening patients to identify high-risk cohorts based on biological and imaging biomarkers could enable more personalized treatment plans and reduce unnecessary risk and costs. Risk factors are well-characterized from trials, and anti-amyloid therapy is increasingly recommended. There is hence a growing need for pretreatment risk stratification tools. Clinically useful tools require individualized predictive models, but existing risk modeling is limited to pharmacokinetic analyses and frameworks using a small handful of pre-specified factors. None integrate multimodal pre-treatment data, use foundation models, or show validation in real-world cohorts. We propose a framework combining multimodal foundation model architectures (imaging, tabular, vision-language) with ARIA-specific design choices, including prediction heads separating ARIA-E and ARIA-H, severity grading, time-to-event modeling, and APOE -stratified monitoring decisions, validated through real-world hospital data in patients receiving ATT.
Authors
- Lawren A. VandeVrede (ORCID: https://orcid.org/0000-0002-8304-6711)
- Andreas M. Rauschecker (ORCID: https://orcid.org/0000-0003-0633-9876)
- Michael F. Romano (ORCID: https://orcid.org/0000-0002-7022-9252)
- Duygu Tosun (ORCID: https://orcid.org/0000-0001-8644-7724)
- Ali Shah Tejani (ORCID: https://orcid.org/0000-0002-6862-5299)
- Jinchi Wei (ORCID: https://orcid.org/0000-0002-4079-949X)
- Evan S Bloch
- Luke Bonham
Institutions
- University of California, San Francisco (US)
- University of California, Berkeley (US)
Publication Details
- Journal
- npj Dementia
- Published
- 2026-11-13
- DOI
- https://doi.org/10.1038/s44400-026-00138-z
- Primary Topic
- Alzheimer's disease research and treatments
- Type
- article
- Field-Weighted Citation Impact
- 0.00