Hemodynamic Digital Twins: A Physics-Informed Artificial Intelligence Approach for Critical Care Prognostics

Abstract

The contemporary Intensive Care Unit (ICU) suffers from reactive monitoring and high false-alarm rates due to frequent physical sensor disconnections. Furthermore, traditional AI forecasting fails in critical care due to autoregressive drift over long horizons and a lack of physiological explainability. To solve these critical barriers, this thesis proposes an end-to-end Medical Digital Twin (MDT) powered by a novel hybrid architecture. 1. Edge Resilience (MambaTS): We engineer a continuous-time Selective State Space Model featuring a Missingness-Aware Gate. By intercepting the Softplus activation to force a Zero-Order Hold identity mapping, the model mathematically freezes its memory during sensor dropouts, preventing state corruption while using surviving sensors to deduce missing data in O(N) linear time. 2. Cloud Forecasting (SST-KODA): To eliminate autoregressive drift, we introduce a Spatio-Temporal Koopman architecture. It bifurcates fast and slow frequencies before lifting chaotic dynamics into a linear Hilbert space, stabilized by a Kalman-inspired Data Assimilation loop for accurate 30-minute predictions. 3. Explainable Physics (PI-DeepONet): We enforce physiological plausibility via an Operator Learning engine that computes the 0D Windkessel equation (Vascular Resistance and Compliance) via Autograd. This achieves cross-patient generalization without online retraining and translates black-box neural weights into true clinical physics. Results: Validation on the VitalDB cohort demonstrates paradigmshifting accuracy. During a simulated 45-second sensor blackout, MambaTS reduced the error to 4.2 mmHg (down from 18.4 mmHg in baselines). SST-KODA achieved an MSE of 5.71 mmHg2 (R2 = 0.9102) for long-horizon forecasting, while PI-DeepONet successfully simulated crosspatient hemodynamics with an MSE of 25.40 mmHg2 (R2 = 0.8412), all while maintaining sub-millisecond edge latency.