Virtual Seminar with Rishi Kamaleswaran: “Forecasting Patient Trajectories: Leveraging Irregularly Sampled, Multimodal Time-Series for Clinical Foundation Models”

As in-patient environments generate vast amounts of heterogeneous data, ranging from continuous physiological telemetry to sparsely sampled medical imaging, there is a profound opportunity to translate this information into actionable clinical foresight. This talk will focus on the development of novel foundation models and transformer architectures designed to handle the temporal complexities of longitudinal hospital data. I will present three recent frameworks developed by our group: Clin-JEPA, which leverages a joint-embedding architecture to forecast complex EHR patient trajectories without the need for per-task fine-tuning; CardioFM, a multimodal foundation model pretrained on over 500,000 hours of physiological data that provides unified representations for advanced tasks like non-invasive hemodynamic trending; and CXR-TFT, which fuses irregular, sparse chest X-rays with high-frequency clinical measurements to predict emerging lung pathologies before they are radiographically visible. Attendees will learn how these sophisticated temporal models overcome data sparsity and modality fragmentation to provide a predictive, “whole-patient” view of hospitalized patients across the acuity spectrum.

Dr. Rishikesan Kamaleswaran is an Associate Professor at the Duke University School of Medicine, holding a primary appointment in the Department of Surgery and secondary appointments in Anesthesiology, Biomedical Engineering, Electrical and Computer Engineering, and Biostatistics and Bioinformatics. He also serves as the Director of Data Science and the Director of the Data Science and Informatics Module for the Duke Clinical & Translational Science Institute. Dr. Kamaleswaran earned his Ph.D. in Computer Science from the University of Ontario Institute of Technology. His biomedical data science research program develops advanced machine learning architectures, multi-modal clinical data fusion, and time-series foundation models for critical care and high-acuity medicine. He serves as Principal Investigator on several major NIH R01 projects and is a Fellow of the Association for Clinical and Translational Science.