Fused Imaging and Large Language Foundational Model for 90-day Stroke Outcome Prediction

The advent of endovascular therapy (EVT) has revolutionized the treatment of acute ischemic strokes (AIS) caused by large vessel occlusions, affecting up to 700,000 Americans annually. However, despite this significant advance, the current approach to EVT patient selection, primarily based on neuroimaging, inadequately addresses the full spectrum of stroke pathophysiology and patient-centered aspects of care, especially long-term functional outcomes. This shortfall is evident in the lack of integration of essential data from clinical notes, which include detailed patient medical history, symptomatology, and lifestyle factors pre-stroke. Notably, there has been limited research on how a more comprehensive approach to patient data could enhance EVT patient selection. Addressing this gap would significantly advance our ability to predict long-term outcomes in AIS patients, with and without treatment, improving EVT patient selection and ultimately clinical outcomes. This proposal aims to address a vital gap in stroke patient care by moving beyond traditional imaging-centric methods to develop a Fused Imaging and Large Language Foundational Model (FILLM). Leveraging advancements in deep learning and large language models, FILLM is a significant innovation that integrates unstructured clinical notes, imaging, and structured tabular clinical data, revolutionizing the prediction of long- term outcomes for stroke patients. The project's three primary objectives include developing and testing FILLM to predict 90-day outcomes in AIS patients, assessing its generalizability and bias reduction capabilities, and enabling optimal collaboration between the FILLM model and clinicians. The training and testing of the FILLM model will use retrospective data from multiple previous clinical trials, as well as registry data from Lausanne, Johns Hopkins, Stanford University hospitals. A key component of this project is the comprehensive career development plan during the K99 phase, which includes mentorship from experts in radiology, neurology, and computer science. This plan involves hands-on training in multi-modal deep learning AI, active engagement in the clinical application of FILLM, and immersive learning experiences through clinical shadowing and academic coursework. These activities, designed to foster essential skills for independent research in the R00 phase, align closely with the research goals, ensuring a holistic approach to both scientific inquiry and career development. Upon successful completion, this project will significantly better our ability to triage EVT patients using data- driven approaches, promoting a more personalized approach to patient care. The NIH's investment in this project will not only bridge a critical gap in stroke treatment but also facilitate my progression to an independent researcher, ready to spearhead future innovations in stroke research.