Foundation Models for Human Functional Neuroimaging with Applications to Psychiatric Disorders

Stanford University, Stanford CA

Investigators

Abstract

Project Abstract Our project aims to harness the transformative potential of Foundation Models (FMs) in the fields of clinical neuroscience and psychiatry. In recent years, a successful new paradigm for building artificial intelligence systems has emerged: train a single model on a vast amount of heterogeneous, unlabeled data and adapt it to various applications. While such FMs have demonstrated unprecedented capabilities in fields as diverse as natural language processing and medicine, their potential for decoding the complexities of human functional brain imaging data remains largely untapped. Crucially, conventional data analysis methods for functional brain imaging (fMRI) are hampered by the need for large, labeled datasets, their inability to capture the intricate spatiotemporal dynamics associated with psychopathology, and the recurrent issues of heterogeneous data and class imbalance in clinical brain imaging studies. To address these challenges, we propose to develop an integrated framework that combines FMs with Explainable Artificial Intelligence (XAI) techniques. This framework will be used to analyze large open-source human brain imaging datasets and phenotypic data across multiple psychiatric disorders. Moreover, our innovative FMs will provide a robust framework for effectively integrating heterogeneous datasets and addressing the recurrent issue of class imbalance in clinical brain imaging studies. Building on our highly promising preliminary results, our team, working with the Stanford Center for Foundation Model Research – a leading institute dedicated to FM research – is uniquely positioned to achieve the following research objectives: First, we will develop FMs specifically designed for large, open- source task-free fMRI datasets such as the HCP, Lifespan Connectome, Human Connectomes Related to Human Disease, and ABCD, among others. Second, we will fine-tune FMs to investigate multiple highly prevalent psychiatric disorders and predict clinical symptom severity. Lastly, we will identify personalized neurobiological features and associated brain networks underlying these disorders. Our FMs will obviate the need for large, labeled datasets, exhibit robustness against class imbalance, and generalize to heterogeneous data. Our approach is theoretically grounded, building upon unifying models of psychopathology that emphasize the role of key control brain networks in multiple psychiatric disorders. By pioneering state-of-the-art FMs tailored for functional brain imaging, our work aims to equip researchers with a transformative toolset for investigating the neurobiology of a range of psychiatric disorders within a unified framework. This will lay the groundwork for more precise diagnostic and treatment strategies in psychiatry. Moreover, the open sharing of FMs and associated data analysis code will catalyze future research, significantly amplifying the impact and scope of our work.