Combining machine learning and physiological network models for sepsis prediction

As the most extreme course of an infectious disease, sepsis poses a serious health threat, with a high mortality rate and frequent long-term consequences for survivors. Despite its enormous burden on global healthcare and ongoing research efforts, early sepsis onset prediction remains challenging due to the complex nature of its pathophysiology. Current approaches face a fundamental trade-off: data-driven machine learning models achieve strong performance but lack interpretability, while biologically inspired models provide mechanistic insights but have limited clinical validation. In this study, we propose the Latent Dynamics Model, a hybrid machine learning approach that integrates a functional model of coupled oscillators representing organ- and immune-cell populations and their interactions. Here, the model parameters encode physiological conditions and allow for an interpretable differentiation between healthy and pathological states. By projecting high-dimensional patient data into the low-dimensional parameter space of the functional model, machine-learned trajectories through this space allow the prediction of critical organ system states and simultaneously offer interpretability beyond plain risk estimates. The proposed method is trained and evaluated on real intensive care patients, achieving competitive AUROC/AUPRC performance on a retrospective MIMIC-IV cohort. Additional qualitative analysis reveals that the learned trajectories exhibit clinically plausible patterns of deterioration, recovery, and stability. We demonstrate that a physiological network model can be embedded within a deep learning architecture without compromising predictive performance while providing an interpretable latent structure for sepsis onset prediction.