State-Space Fusion of Physiologic, Imaging, and Molecular Data for Early Anastomotic Failure Detection

Abstract

Clinical risk prediction in surgery is still dominated by structured variables that appear only after charting, aggregation, and manual abstraction. That timing constraint matters because physiologic deterioration often begins before a formal entry exists in the electronic health record. In perioperative care, the consequence is a recurrent mismatch between when a patient starts to deviate from a stable trajectory and when decision support systems are finally able to measure that deviation. This paper develops a technical framework for multimodal earlier-than-EHR risk signaling in perioperative anastomotic failure, a setting in which delays in recognition can change both salvage options and downstream morbidity. The proposed formulation treats earlier-than-EHR information not as a vague notion of temporal advantage, but as an estimable property of raw data streams whose acquisition time precedes structured chart availability. Continuous physiologic waveforms, intraoperative perfusion imaging, molecular compositional profiles, device telemetry, and sparse structured records are fused through a continuous-time latent state model linked to a hazard process for failure. The framework addresses asynchronous sampling, variable data quality, missing-not-at-random observations, label delay, and deployment constraints under finite alert budgets. It also distinguishes prediction performance from lead-time utility by defining metrics that reward accurate alarms only when they arrive early enough to matter clinically. The result is a unified research design for modeling, training, and evaluating pre-charting surgical risk signals. Rather than assuming that more modalities automatically improve prediction, the paper argues that the primary technical advantage of multimodal systems lies in recovering temporally proximal evidence before it is compressed into late EHR abstractions.