Interpretable machine learning based comorbidity specific mortality risk score for bloodstream infections

Bloodstream infections (BSI) are a leading cause of sepsis, necessitating early risk stratification during the critical window for effective management. We aimed to develop a model that leverages early-phase clinical data to predict 28-day mortality by integrating longitudinal trends and addressing comorbidity-driven heterogeneity.

The BSI Heterogeneity Score (BHScore) was developed using machine learning on longitudinal clinical data (first 7 days post-culture) from 2524 BSI patients at Xijing Hospital. The model uses interpretable methods to establish comorbidity-stratified thresholds (global, renal disease, liver disease, metastatic malignancy) to enhance prediction accuracy.

The BHScore demonstrated superior discriminatory performance (area under the receiver operating characteristic curve: 0.81-0.91) and temporal stability compared to established scores including the Sequential Organ Failure Assessment (SOFA), representing an improvement of 10% to 25%. Our analysis revealed that coagulation biomarkers have greater prognostic significance in the malignancy subgroup, inflammatory thresholds are more sensitive in the liver disease subgroup, and urea levels exhibit U-shaped mortality curves in the renal disease subgroup. To support clinical application, a freely accessible web tool has been developed.

The BHScore enables the early identification of high-risk BSI patients with diverse comorbidities using straightforward clinical indicators, facilitating timely and targeted interventions to reduce mortality.