Multimodality and Longitudinal Artificial Intelligence for Diagnosis and Prognosis in Hepatic Steatosis - Project Summary Metabolic dysfunction-associated steatotic liver disease (MASLD) affects approximately one-third of adults in the United States. If left untreated, patients may develop more severe disease including steatohepatitis, cirrhosis or hepatocellular carcinoma. There are significant challenges to diagnosis, prognostication, and treatment. In MASLD, clinical biomarkers are crucial for guiding patient care, risk assessment, and outcome prediction in medicine and include elevated alkaline phosphatase in the setting of biliary obstruction or hepatic loss of signal on out-of-phase MR imaging in the setting of hepatic steatosis. However, predictive models have significant limitations: (1) it is difficult to extrapolate risk models across tests or integrate multimodal data into a single model and (2) the models use data only at a particular moment in time. This is the approach that most machine learning models take and represents a significant limitation, as it is diametrically opposed to the approach that radiologists take, which involves incorporating longitudinal data via comparison to prior imaging and leveraging the strengths of multiple imaging modalities in conjunction with clinical data. This limited focus is a missed opportunity, with the electronic health record (EHR) containing vast amounts of unused longitudinal and multimodal quantitative and imaging data that can guide patient care. The investigators propose developing deep learning tools to predict patient risk, synthetic clinical biomarkers, and normative data from multimodal and longitudinal health data. The deep learning tools will be trustworthy and reliable, providing appropriate error estimates and confidence intervals in the context of missing data. Building on our experience in deep learning for medical imaging and multimodal satellite imagery data, in Aim 1 the investigators propose to develop novel deep learning techniques that utilize multimodal and longitudinal health data to determine likelihood of MASLD. In Aim 2, the investigators propose to determine the performance of these models in an out-of-domain cohort of patients and compare with validated MASLD identification algorithms. The ultimate impact of this work will be new and highly accurate deep learning models that predict outcomes related to cardiovascular disease, diabetes, and liver disease and result in interpretable and actionable criteria for physicians.
