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

Area of Science:

• Clinical Medicine
• Infectious Diseases
• Medical Informatics

Background:

• Bloodstream infections (BSI) are a major cause of sepsis, requiring early risk assessment for effective treatment.
• Existing risk stratification tools may not fully capture the complexity of BSI patients with diverse comorbidities.

Purpose of the Study:

• To develop and validate a predictive model for 28-day mortality in BSI patients.
• To integrate longitudinal clinical data and account for comorbidity-driven heterogeneity in risk prediction.
• To improve early risk stratification for timely intervention in BSI-related sepsis.

Main Methods:

• Developed the BSI Heterogeneity Score (BHScore) using machine learning on longitudinal clinical data (first 7 days) from 2524 BSI patients.
• Employed interpretable methods to establish comorbidity-stratified thresholds for renal disease, liver disease, and metastatic malignancy.
• Utilized data from Xijing Hospital for model development and validation.

Main Results:

• BHScore demonstrated superior discriminatory performance (AUC: 0.81-0.91) and temporal stability compared to SOFA score, improving prediction by 10-25%.
• Identified subgroup-specific prognostic indicators: coagulation biomarkers (malignancy), inflammatory thresholds (liver disease), and urea levels (renal disease).
• A freely accessible web tool was developed to support clinical application of the BHScore.

Conclusions:

• The BHScore effectively identifies high-risk BSI patients with various comorbidities using simple clinical indicators.
• Facilitates early and targeted interventions, potentially reducing mortality in sepsis patients.
• Highlights the importance of considering comorbidity heterogeneity in BSI risk prediction.