Home › Research › Brain scans and imaging
Current gaps in downstream applications of fMRI foundation models: lessons learned from BrainLM & BrainHarmonix on Prevent-AD
Wang H, Pichet Binette A, Sainath P, Farrugia N, Villeneuve S, Bellec L.. medRxiv, 2026-10-26
Topic: Brain imaging, connectivity and assessment
Foundation models have driven major advancements in the language and vision domains, as seen in the popularity of Artificial Intelligence (AI) tools, and a growing number of studies are bringing the approach to human brain imaging. Whether these models are ready for independent researchers to use on clinical datasets has yet to be examined. We evaluated two recent fMRI foundation models, BrainLM and BrainHarmonix, on PreventAD, a dataset on adults with a family history of Alzheimer’s disease. We compared features from frozen and fine-tuned models against two baselines: functional connectivity and a dummy classifier that always predicts the most common class. The sample includes 343 subjects who passed quality control, with a subset of 218 who had amyloid PET imaging. We tested four prediction targets: sex, age, progression to mild cognitive impairment (MCI), and β-amyloid burden. Both models required substantial changes to the released code. BrainLM’s own data preparation workflow did not produce stable fine-tuning, which we achieved only after switching to a standard workflow implemented with the library Nilearn. BrainHarmonix’s documentation did not specify how to normalize new data, so we resolved this by testing different alternatives. For sex, functional connectivity outperformed all foundation model features. For age group, the best foundation model feature, taken from BrainHarmonix’s structural (T1-weighted) encoder, slightly exceeded functional connectivity, but the difference was not significant. For MCI and β-amyloid, no feature exceeded the dummy classifier in accuracy, and the area under the curve (AUC) stayed near chance for all features. For both clinical targets, the most precise feature came from fine-tuned BrainLM, meaning a larger share of its positive predictions were correct than for functional connectivity, but neither difference reached statistical significance. These results suggested that the strong transfer to new tasks seen in large language models cannot yet be assumed for fMRI foundation models, and that both models required code adaptation that would preclude their streamlined adoption by independent groups. We recommend that future studies provide standardized data workflows and always report functional connectivity and dummy classifier baselines. For clinical targets with strong class imbalance, studies should report several metrics, including precision.
Cite this study APA, Vancouver, BibTeX · download .bib or .ris
APA 7th
Wang, H., Pichet Binette, A., Sainath, P., Farrugia, N., Villeneuve, S., & Bellec, L. (2026). Current gaps in downstream applications of fMRI foundation models: lessons learned from BrainLM & BrainHarmonix on Prevent-AD. medRxiv. https://doi.org/10.64898/2026.10.22.26366227
Vancouver
Wang H, Pichet Binette A, Sainath P, Farrugia N, Villeneuve S, Bellec L. Current gaps in downstream applications of fMRI foundation models: lessons learned from BrainLM & BrainHarmonix on Prevent-AD. medRxiv. 2026. doi:10.64898/2026.10.22.26366227.
BibTeX
@article{wang2026-51eea5af,
author = {Wang, H. and Pichet Binette, A. and Sainath, P. and Farrugia, N. and Villeneuve, S. and Bellec, L.},
title = {{Current gaps in downstream applications of fMRI foundation models: lessons learned from BrainLM \& BrainHarmonix on Prevent-AD}},
journal = {medRxiv},
year = {2026},
doi = {10.64898/2026.10.22.26366227},
url = {https://doi.org/10.64898/2026.10.22.26366227}
}
Built from the record alzhub holds for this paper, not looked up. Volume, issue and page numbers are not held in the record. Check the citation against the original before you use it.
Key findings
- For age group, the best foundation model feature, taken from BrainHarmonix’s structural (T1-weighted) encoder, slightly exceeded functional connectivity, but the difference was not significant.
- These results suggested that the strong transfer to new tasks seen in large language models cannot yet be assumed for fMRI foundation models, and that both models required code adaptation that would preclude their streamlined adoption by independent groups.
New to this subject? Start with our plain-English guide: Brain scans & tests →
Study details
- Participants343
- Evidence strengthLow
How this study scored 16/100
Preprint — not yet peer-reviewed. It scored 29 on the same rubric, but non-peer-reviewed sources are capped at 16, so it shows as 16. Scored zero for study design, effect reported and randomised / blinded because the record does not state it. We have no data on times cited, citations per year, impact for its field and year, cited by studies here, still being cited, senior author's track record, institution's track record and multi-centre collaboration, so those were left out: it earned 18 of the 62 points we could judge it on, which is 29 out of 100.
| Component | Points | Why |
|---|---|---|
| Journal tier | 2 / 10 | Tier 4 — not a peer-reviewed journal (medRxiv) |
| Journal's citation record | 0 / 4 | Not a journal article, so the journal's record does not count |
| Study design | 0 / 12 | No study design recorded — scores zero, not guessed (not stated in the record — scored 0) |
| Strength of evidence | 1 / 5 | Rated “low” in our summary |
| Peer-review validation | 4 / 10 | Preprint — not yet peer-reviewed. Hosted on a preprint server. |
| How recent | 5 / 5 | 2026 — within 1 year(s) of the newest study here |
| Sample size | 2 / 4 | 343 participants |
| Effect reported | 0 / 3 | No effect size reported — scores zero (not stated in the record — scored 0) |
| Studied in people | 3 / 3 | Studied in people (participant count stated) |
| Randomised / blinded | 0 / 2 | No randomisation or blinding stated — scores zero (not stated in the record — scored 0) |
| Times cited | — / 7 | No citation count (not found in OpenAlex) — left out of this score |
| Citations per year | — / 5 | No citation count or no year — left out of this score |
| Impact for its field and year | — / 6 | No field-weighted citation impact (OpenAlex has none yet) — left out of this score |
| Cited by studies here | — / 6 | Not yet checked against this library's citation graph — left out of this score |
| Still being cited | — / 4 | No yearly citation counts — left out of this score |
| Senior author's track record | — / 4 | Senior author's record not available — left out of this score |
| Institution's track record | — / 3 | No institution listed in OpenAlex — left out of this score |
| Multi-centre collaboration | — / 3 | Institutions not listed — left out of this score |
| Indexed in PubMed | 1 / 2 | Not in PubMed, but has a DOI |
| Free to read | 0 / 2 | Behind a paywall — scores zero |
| Points | 18 / 62 | Scored on 12 of 20 parts. We have no data for the rest, so they are left out: 18 of the 62 points we could judge it on is 29 out of 100. |
| Total | 29 / 100 | Capped to 16 because this is not a peer-reviewed article. |
What each of these 20 components measures, and what this score is not →
Read the original
Free copy unknown: no free-copy check has an answer for this record yet. That does not mean it is paywalled.
Related studies in Brain scans and imaging
Also tagged “convolutional”
205 studies in the library carry the topic tag “convolutional”; these are this study's neighbours in a fixed, alphabetical-by-hash tour of them, not the closest in content. All 205 tagged studies.