Current gaps in downstream applications of fMRI foundation models: lessons learned from BrainLM & BrainHarmonix on Prevent-AD

The human brain is a vast, humming network of billions of cells that communicate through electrical signals. To understand how this network works, scientists often use a powerful camera called functional magnetic resonance imaging, or fMRI. This machine does not take a static photograph; instead, it records a movie of the brain as it rests, capturing the subtle waves of activity that ripple through different regions. By watching how these regions rise and fall in sync, researchers can map the brain's functional connections, much like tracing the flow of traffic on a city map to see which neighborhoods are most closely linked. For years, the standard way to analyze these maps has been to measure the strength of the connections between every pair of regions. This method, known as functional connectivity, has been the workhorse for predicting things like a person's age or sex, and for trying to spot early signs of diseases like Alzheimer's.

Recently, a new wave of artificial intelligence has swept through science, promising to revolutionize how we understand complex data. In fields like language and vision, "foundation models" have become famous. These are massive computer programs trained on enormous amounts of data to learn the underlying patterns of the world without being told specific answers. They learn to predict what comes next in a sentence or what an image looks like, building a deep, flexible understanding that can be applied to new, smaller tasks. Researchers in brain science have been eager to try this approach, hoping to train similar giant models on thousands of brain scans so they could learn the general "grammar" of brain activity. The hope was that these models would become powerful tools that any doctor or scientist could use to diagnose diseases or predict who might develop memory problems, even with very little data of their own.

A team of researchers set out to test whether this promise was real for brain imaging. They chose two of the newest and most advanced foundation models for fMRI, named BrainLM and BrainHarmonix, and put them to the test on a challenging real-world dataset. This dataset came from the PREVENT-AD study, which follows hundreds of healthy older adults who have a family history of Alzheimer's disease. The goal was to see if these new artificial intelligence models could outperform the old, trusted methods of functional connectivity when trying to predict four specific things: a person's sex, their age group, whether they would develop mild cognitive impairment in the future, and whether they had a specific type of protein buildup in their brain that signals Alzheimer's. The researchers wanted to know if these models were ready for independent scientists to use, or if they were still too difficult to apply outside the labs that created them.

The journey began with a significant hurdle: simply getting the models to work on new data was far from straightforward. When the team tried to use the software code provided by the creators of BrainLM, they found it did not work with their data. The code relied on a custom, home-built way of cleaning the brain scans that was incompatible with standard tools used by most neuroscientists. The researchers had to rewrite the data processing steps, switching to a widely accepted, standard method used by the broader community. Only after making these substantial changes did the model begin to learn properly. The second model, BrainHarmonix, presented a different kind of problem. Its documentation did not explain how to prepare new data for the model, leaving the team to guess the correct way to normalize the signals. They had to test several different approaches just to find one that produced stable results. These initial struggles highlighted a major gap: while the models were powerful, they were not yet packaged in a way that allowed independent researchers to use them easily without significant technical troubleshooting.

Once the models were running, the team compared their predictions against the standard method of functional connectivity and a simple "dummy" classifier that just guessed the most common answer every time. For predicting sex, the old method of functional connectivity was actually the winner. It correctly identified the sex of the participants more often than any of the features extracted from the new foundation models. For predicting age groups, one feature from the BrainHarmonix model, which was derived from structural brain images rather than the functional activity, performed slightly better than the standard method, but the difference was so small that it could easily have happened by chance. The results were even more sobering for the clinical targets. When trying to predict who would develop mild cognitive impairment or who had the Alzheimer's-related protein buildup, none of the foundation models could beat the simple dummy classifier in terms of overall accuracy. This was largely because the groups of people with these conditions were very small compared to the healthy group, making it hard for any model to find a signal without just guessing the majority.

However, the story was not entirely one of failure. When the researchers looked more closely at the specific details of the predictions, they found a subtle hint of promise. While the models were not accurate enough to be used as a diagnostic tool, some of the features from the foundation models were better at identifying the few people who actually had the disease than the standard method was. In other words, when the model did make a positive prediction, it was slightly more likely to be correct, even though it missed many cases. This suggests that the models might be capturing a different, perhaps more subtle, kind of signal in the brain data that the standard method misses. Yet, because these improvements were not statistically significant, the researchers could not claim that the models had definitively solved the problem. The findings suggest that the powerful transfer of knowledge seen in language models has not yet been replicated in brain imaging.

The study concludes that these foundation models are not yet ready for independent researchers to adopt for clinical use. They require too much technical adaptation, and they have not yet proven they can outperform the established, simpler methods that have been used for decades. The authors argue that for these models to become truly useful, the field needs standardized software tools that make them easy to install and run, and future studies must always compare new models against the standard functional connectivity baseline. Until then, the dream of a universal brain model that can instantly diagnose Alzheimer's or predict memory decline remains just that—a dream, waiting for more rigorous testing and better engineering to become a reality.