This project develops a scalable ICU Digital Twin framework to model patient trajectories and predict clinical deterioration using electronic health record data from MIMIC-IV. ICU environments are characterized by rapidly evolving patient conditions, irregularly sampled data, and complex multi-modal signals. Those challenges render traditional models insufficient for both outcome prediction and treatment strategy evaluation. To address these challenges, we propose a two -component modeling framework encompassing patient trajectory modeling and counterfactual simulation. The first component employs Neural Controlled Differential Equations (Neural CDEs) to learn continuous-time representations of patient trajectories, preserving temporal structure without requiring discretization or imputation. Building on this latent representation, a Mamba-based selective state-space model with multimodal fusion captures long-range dependencies across clinical signals and events, enabling accurate prediction of outcomes including mortality and sepsis onset. The second component explores probabilistic and event -based approaches, including Hawkes processes and state-space models, to support counterfactual analysis. By integrating causal disentanglement with event-based modeling, the system simulates how intervention timing and intensity influence patient trajectories. For instance, the model can compare early versus delayed treatment scenarios and quantify their downstream effects on physiological signals such as lactate progression. The overall system is structured as a modular pipeline comprising a data layer for cohort construction and preprocessing, a modeling layer for continuous-time representation and prediction, and a simulation layer for counterfactual rollout and trajectory c omparison. Evaluation results demonstrate strong alignment between predicted and observed trajectories, as well as the ability to generate clinically meaningful counterfactual scenarios. This framework establishes the foundation for an ICU Digital Twin, a dynamic virtual patient representation capable of anticipating clinical deterioration and evaluating alternative treatment strategies. Future work will incorporate clinician -facing tools for real-time counterfactual exploration. Broadly, this project advances machine learning in critical care from static risk prediction toward actionable decision support, highlighting the promise of digital twin systems for personalized, data-driven clinical practice. Introduction and Problem Statement Intensive Care Units (ICUs) are among the most complex and high-stakes environments in healthcare, where clinicians must make rapid decisions based on continuously evolving patient conditions. Patient states can deteriorate quickly, and clinical decisions, such as when to administer medications or initiate interventions, must often be made under uncertainty. At the same time, modern ICUs generate large volumes of electronic health record (EHR) data, including vital signs, laboratory results, medications, and clinical events. These data are inherently high -dimensional, irregularly sampled, and multi -modal, presenting significant challenges for traditional data-driven modeling approaches. Existing machine learning methods in healthcare have primarily focused on risk prediction tasks, such as forecasting mortality or the onset of conditions like sepsis. While these models can provide valuable early warning signals, they are limited in the
Mentor: David Banks
Project poster (PDF)
