Photo: Phil Boorman/Getty Images
Using a form of machine learning called self-supervised learning, Mass General Brigham researchers have created a new predictive artificial intelligence model, which they say could help generate insights from sparser medical datasets.
Called BrainIAC, the AI tool can perform specific medical tasks and learn from the process. In their report, published Thursday in Nature Neuroscience, researchers show how the SSL model can predict outcomes where training data is scarce and hold a high degree of accuracy despite task complexity.
Trained on upward of 49,000 diverse brain MRI scans to use key neurological health indicators, such as age factors and tumor mutations, to predict dementia risk, brain cancer survival and other diseases, the model could also adapt well to other real-world settings where annotated medical datasets are not always readily available, say MGB researchers.
WHY IT MATTERS
Mass General Brigham developed the self-supervised foundational model to identify inherent features from brain MRI datasets, finding a high degree of accuracy for the tasks they trained it to perform.
In real-world settings, medical datasets are not always readily available. Brain MRI images from different institutions can vary in appearance and also due to medical needs, such as neurology versus oncology care, making it challenging for AI frameworks to learn similar information from them.
"On one end of the spectrum, MRI sequence classification and tumor segmentation are straightforward for trained clinicians and, on the other end of the spectrum, time-to-stroke prediction, genomic subtyping and survival prediction are very challenging based on imaging alone," researchers said.
BrainIAC could successfully generalize its learnings across healthy and abnormal images and subsequently apply them to both relatively straightforward tasks, such as classifying MRI scan types, and very challenging tasks, such as detecting brain tumor mutation types, they explained.
Researchers used seven "pretext tasks" to enable the AI to learn from raw, unannotated data at scale and make classifications and determinations. They then ran it through a series of tests and found that BrainIAC outperformed three more conventional, task-specific AI frameworks.
"Our findings demonstrate BrainIAC’s versatility and ability to adapt to several clinical settings with extremely limited training data (as few as single samples), providing a usable foundation to accelerate computational brain imaging analysis research," they said.
In addition to predicting early dementia through mild cognitive impairment classification, other tasks, like time-to-stroke prediction, hold additional clinical uses.
"Time-since-stroke-onset prediction can allow clinicians to make evidence-based decisions for time-sensitive treatments, enabling optimal intervention selection for stroke patients who might otherwise be excluded due to uncertain onset time," the researchers said.
Because the AI model can predict outcomes when training data is scarce or when task complexity is high, the AI could also be adapted to other healthcare applications. The researchers said that further research is needed to test the framework on additional brain imaging methods and larger datasets.
THE LARGER TREND
Much research led by MGB's Artificial Intelligence in Medicine Program focuses on improving speed-to-care for patients and getting ahead of disease.
For instance, a deep learning algorithm called FaceAge could allow clinicians to improve their qualitative assessments and possibly catch diseases sooner, according to MGB oncologist Dr. Raymond Mak, another AIM researcher, who will showcase biomarker technology research at HIMSS26 in Las Vegas next month.
ON THE RECORD
"BrainIAC has the potential to accelerate biomarker discovery, enhance diagnostic tools and speed the adoption of AI in clinical practice," said Dr. Benjamin Kann, a radiation oncologist with MGB and one of the researchers at AIM. "Integrating BrainIAC into imaging protocols could help clinicians better personalize and improve patient care."
Andrea Fox is senior editor of Healthcare IT News.
Email: afox@himss.org
Healthcare IT News is a HIMSS Media publication.
