Researchers at Baylor College of Medicine and Yale University have developed a generative artificial intelligence (AI) model, known as Brain Language Model (BrainLM), to better understand the relationship between brain activity, human behavior, and brain diseases. The findings were presented at ICLR 2024, a leading conference in the field of deep learning AI.
"For a long time we’ve known that brain activity is related to a person’s behavior and to a lot of illnesses like seizures or Parkinson’s," said Dr. Chadi Abdallah, associate professor in the Menninger Department of Psychiatry and Behavioral Sciences at Baylor and co-corresponding author of the paper. He added that traditional data analytical tools had limitations in capturing the dynamics of these activities over time and space.
Generative AI tools can create foundational models independent of specific tasks or patient populations. They act as detectives uncovering hidden patterns within datasets by analyzing data points and their relationships. These models learn the underlying dynamics - how and why things change or evolve - which are then fine-tuned to understand various topics.
The team utilized generative AI to capture how brain activity functions irrespective of any particular disorder or illness. This method can be applied to any population without needing information about their behavior, illness history, or age; it simply requires their brain activity data.
By training the model with 80,000 scans from 40,000 subjects, researchers established BrainLM as a foundational model for understanding how brain activities relate over time. This model can now be used for specific tasks in other studies.
Abdallah explained that BrainLM could potentially reduce costs in clinical trials by selecting individuals most likely to benefit from treatment based on knowledge learned from previous scans.
The first step involved preprocessing signals and removing noise irrelevant to brain activity. The summaries were then input into a machine learning model with some data masked for each individual. The model was tested on different samples including older adults and younger adults, using different scanners. The results showed that BrainLM performed well in various samples and predicted depression, anxiety, and PTSD severity better than other machine learning tools not using generative AI.
Abdallah said, "We found that BrainLM is performing very well. It is predicting brain activity in a new sample that was hidden from it during the training as well as doing well with data from new scanners and new population." He added that they are now working on increasing the training dataset to build a stronger model for assisting patient care.
The researchers plan to use this model in future research to predict illnesses related to the brain. Contributors to this work include Josue Ortega Caro, Antonio Henrique de Oliveira Fonseca, Syed A Rizvi, Matteo Rosati, Christopher Averill, James L Cross, Prateek Mittal, Emanuele Zappala, Rahul Madhav Dhodapkar and David van Dijk who are affiliated with Baylor College of Medicine, Yale University, University of Southern California and Idaho State University. This work was funded by Yale’s Wu Tsai Institute and Baylor’s Beth K. and Stuart Yudofsky Chair in the Neuropsychiatry of Military Post Traumatic Stress Syndrome.
