False hope of a single generalisable AI sepsis prediction model: bias and proposed mitigation strategies for improving performance based on a retrospective multisite cohort study.

Objective: To identify bias in using a single machine learning (ML) sepsis prediction model across multiple hospitals and care locations; evaluate the impact of six different bias mitigation strategies and propose a generic modelling approach for developing best-performing models.

Methods: We developed a baseline ML model to predict sepsis using retrospective data on patients in emergency departments (EDs) and wards across nine hospitals. We set model sensitivity at 70% and determined the number of alerts required to be evaluated (number needed to evaluate (NNE), 95% CI) for each case of true sepsis and the number of hours between the first alert and timestamped outcomes meeting sepsis-3 reference criteria (HTS3). Six bias mitigation models were compared with the baseline model for impact on NNE and HTS3.

Results: Across 969 292 admissions, mean NNE for the baseline model was significantly lower for EDs (6.1 patients, 95% CI 6 to 6.2) than for wards (7.5 patients, 95% CI 7.4 to 7.5). Across all sites, median HTS3 was 20 hours (20-21) for wards vs 5 (5-5) for EDs. Bias mitigation models significantly impacted NNE but not HTS3. Compared with the baseline model, the best-performing models for NNE with reduced interhospital variance were those trained separately on data from ED patients or from ward patients across all sites. These models generated the lowest NNE results for all care locations in seven of nine hospitals.

Conclusions: Implementing a single sepsis prediction model across all sites and care locations within multihospital systems may be unacceptable given large variances in NNE across multiple sites. Bias mitigation methods can identify models demonstrating improved performance across most sites in reducing alert burden but with no impact on the length of the prediction window.