Patient-Referenced and Trajectory Features from Serial ECGs for Haemodynamic Monitoring: An In Silico Study with External Feasibility Analysis

Abstract Objective. Physiological measurements are read against population thresholds, although the within-person change that signals deterioration is smaller than the spread between people. We asked, in simulation, whether serial electrocardiograms read against a patient's own earlier recording carry information about haemodynamic state, and whether the trajectory of that change adds to its current value, in acute haemorrhage. Approach. A nine-dimensional circulatory state-space model generated 600 synthetic subjects in whom haemorrhage reached the waveform only through subject-specific couplings. Results were audited against elapsed time and heart rate, repeated under alternative observation models and with an unknown bleeding onset, and examined in an external feasibility analysis of 77 intraoperative recordings. Main results. Under the generative model, a personal reference raised state discrimination from 0.847 to 0.951 but helped event prediction less than the slope and curvature of the change, which raised five-minute decompensation discrimination from 0.876 to 0.953. Elapsed time, patient characteristics and heart rate reached 0.930 without any electrocardiogram; with onset randomised that shortcut disappeared and the waveform still added +0.084. This division of labour held in all 7 observation models and 24 sampled coupling sets. Intraoperatively, patient-referenced features discriminated concurrent hypotension better (0.683) than absolute features or heart rate, but added nothing distinguishable from zero to the recorded anaesthetic and ventilatory state. Significance. Under the stated assumptions, a personal reference estimates where a patient is and its trajectory where they are going. The external cohort provides preliminary evidence for the portability of the patient-referenced representation but is insufficient to establish clinical performance.