An Artificial Intelligence Foundation Model for Functional Neuroimaging: Personalized Prediction, Treatment Stratification, and Biotype Discovery in Major Depressive Disorder - PROJECT SUMMARY Major depressive disorder (MDD) is a leading cause of disability, with substantial individual and societal costs. The heterogeneity of MDD and the lack of predictive tools for individualized treatment present significant challenges to effective care. This proposal aims to leverage recent advances in foundation models, a type of artificial intelligence (AI) that has demonstrated remarkable success in natural language processing, to develop a neuroimaging-based tool that can aid in prognostication, treatment stratification, and biotype discovery in MDD. Foundation models are pretrained on massive datasets, enabling them to learn generalizable features that can then be adapted to smaller, more specific datasets. This approach is ideally suited for psychiatric neuroimaging, where clinical datasets are scarce; however, non-clinical datasets like the Human Connectome Project and UK Biobank are extensive. I have developed a functional prototype by adapting a transformer architecture to analyze functional magnetic resonance imaging (fMRI) time-series data and training it on the UK Biobank. Preliminary data generated using this prototype indicate strong potential for this approach. Applying this innovative technique to psychiatry holds great promise for advancing the understanding and treatment of MDD. To achieve this, I propose three specific aims. Aim 1: Use pooled fMRI data from individuals with MDD to fine-tune the pretrained model to decode depression severity and uncover MDD biotypes; Aim 2: Use pooled fMRI scans from longitudinal treatment data to fine-tune the pretrained model to predict antidepressant response and identify neural circuits of treatment response; Aim 3: Prospectively evaluate the performance of MRI-based treatment prediction models in a pilot clinical trial. If successful, this work will yield a novel neurocomputational framework for personalized treatment stratification and significantly advance our understanding of MDD neurobiology and heterogeneity. Through this research, training, and expert mentorship, I will gain expertise in: 1) AI foundation models, including transformer architectures and interpretability techniques; 2) applying foundation models to neuroimaging to generate clinically actionable predictions and mechanistic insights; 3) clinical trial design and analysis of longitudinal data; and 4) professional skills for transitioning to independence. The training plan—which includes coursework, workshops, close mentorship, and hands-on research experience—builds on my existing expertise in neuroimaging, network neuroscience, and clinical psychiatry. Stanford University offers an exceptional environment with access to cutting-edge computational resources, neuroimaging facilities, and a vibrant community of AI experts and clinician-scientists. In sum, through the K23 award, the proposed research, training, mentorship, and pilot data will enable me to successfully compete for independent research funding and establish a high-impact patient-oriented research program in neurocomputational psychiatry at the intersection of AI, neuroimaging, and precision treatment.
