Answers are grounded in this paper and cite their sources.
Early-warning models for sepsis are usually evaluated on clean data, yet bedside data can be wrong in ways a model cannot see, such as a temperature in the wrong units, a monitor artefact or a frozen feed, and the model can remain confident. We present SepsisShield, an end-to-end system that pairs a calibrated sepsis-risk ensemble of five LightGBM models with an independent input-integrity layer that grades the inputs of every patient-hour as HIGH, REDUCED or LOW trust and turns the grade into an action: show, warn or withhold. Only input evidence can withhold a prediction. On 8,068 held-out patients from the PhysioNet/Computing in Cardiology 2019 data, the model reached an AUROC of 0.852 (95% CI 0.840–0.865) and alerted 53.9% of septic patients at least 6 h before onset. Across four simulated accidental fault types, the system flagged or withheld 95.4% (6,481/6,790) of alert decisions that a fault had flipped to a wrong answer (validation replication 94.7%) while withholding 0.46% of clean patient-hours. A component ablation showed that ensembling and calibration alone caught none of these failures and ensemble disagreement caught 4.0%. When only 1% of clean hours can be reviewed, ranking by input trust caught 47.3% of harmful decisions, against 4.4% for model uncertainty; an offline, validation-fitted combination of signals caught 82.2%. Input-integrity evidence thus identified fault-induced unsafe decisions that model-side confidence largely missed, while withholding fewer than 0.5% of clean predictions. Performance remains limited for plausible deliberate edits, frozen feeds and transfer to an unseen hospital.
No takes yet. Share an insight, caveat, or question.
Israt Jahan Aunika (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: