AI-Based Predictive Models for Sepsis

Abstract

Sepsis -- life-threatening organ dysfunction caused by a dysregulated host response to infection -- affects 49 million people annually with a mortality rate of 20-30%, and each hour of delayed treatment increases mortality by 4-8%. Early prediction enabling pre-emptive intervention is therefore critical, yet current clinical scoring systems (SOFA, qSOFA, SIRS) detect sepsis only after organ dysfunction is established. AI-based predictive models analysing high-frequency electronic health record (EHR) data can identify subtle pre-septic physiological patterns hours before clinical recognition. This study developed and benchmarked six sepsis prediction architectures -- SIRS criteria, qSOFA, SOFA trend analysis, gradient-boosted trees (XGBoost), recurrent neural network (LSTM), and a proposed Temporal Graph Attention Network for Sepsis (TGA-Sepsis) integrating temporal vital sign dynamics, laboratory trajectories, medication context, and inter-organ dependency modelling -- across two large ICU cohorts: MIMIC-IV (n = 52,846 ICU stays, 11.4% sepsis) for development and eICU-CRD (n = 34,628 stays, 9.8% sepsis) for external validation. TGA-Sepsis achieved the highest AUROC for sepsis onset prediction at 6 hours before clinical recognition: 0.928 +- 0.004 (MIMIC-IV) and 0.904 +- 0.006 (eICU-CRD), significantly outperforming LSTM (0.892; p < 0.001), XGBoost (0.876; p < 0.001), and SOFA trend (0.812; p < 0.001). At the clinically optimal operating point (sensitivity 90%), TGA-Sepsis achieved specificity 82.4% with PPV 48.6% -- meaning fewer than 3 alerts per true sepsis case -- and median prediction lead time of 6.4 hours (IQR 3.8-9.2) before Sepsis-3 criteria were met. The temporal graph attention mechanism learned clinically interpretable inter-organ deterioration cascades: renal-hepatic coupling preceded cardiovascular collapse in 72.4% of sepsis cases, providing actionable organ-specific early warning. These results establish graph-based temporal modelling as the state-of-the-art for sepsis prediction and provide an externally validated, interpretable framework for clinical deployment.