Predictive model for daily risk alerts in sepsis patients in the ICU: visualization and clinical analysis of risk indicators.

Summary

A two-stage Transformer time-series model trained on eICU data achieved AUC 0.87 on admission, improving to 0.92 by day 5. External validation showed accuracy 81.8% (AUC 0.73) on Chinese sepsis data and 76.56% (AUC 0.84) on MIMIC-IV-3.1, with SHAP-based temporal heatmaps linking predictions to interpretable biomarkers.

Key Findings

• Transformer-based two-stage model achieved AUC 0.87 (±0.021) on admission day, rising to 0.92 (±0.009) by day 5.
• External validation: 81.8% accuracy (AUC 0.73) on Chinese sepsis cohort and 76.56% accuracy (AUC 0.84) on MIMIC-IV-3.1.
• SHAP-derived temporal heatmaps revealed mortality-associated feature dynamics, enhancing interpretability for clinicians.

Clinical Implications

Supports deployment of dynamic, interpretable risk dashboards to prioritize care, trigger earlier diagnostics or therapies, and allocate ICU resources; prospective implementation studies are warranted.

Limitations

• Retrospective training/validation; prospective impact not yet proven
• Potential dataset shift and variable availability differences across sites

Future Directions

Prospective, multi-center implementation trials to test workflow integration, clinical impact on time-to-intervention and mortality, and adaptive recalibration for dataset shift.