Integrating Nonindividual Patient Features in Machine Learning Models of Hospital-Onset Bacteremia.

Area of Science:

• Infectious Disease Epidemiology
• Machine Learning in Healthcare
• Hospital-Acquired Infections

Background:

• Hospital-onset bacteremia and fungemia (HOB) are significant complications of inpatient care.
• Predicting and preventing HOB is crucial for patient safety and outcomes.

Purpose of the Study:

• To evaluate the utility of non-individual patient features in machine learning models for predicting HOB.
• To assess the contribution of patient interactions and healthcare worker (HCW) exposures to HOB risk.

Main Methods:

• A prognostic study utilizing electronic health records from an academic hospital.
• Development of gradient boosting models, including predictive and causal approaches.
• Engineering of non-patient features, such as room occupancy history and HCW contact rates.

Main Results:

• Models incorporating non-patient features demonstrated improved predictive performance (AUROC, 0.88; AUPRC, 0.20) compared to patient-only models (AUROC, 0.85; AUPRC, 0.13).
• Specific non-patient factors, including prior room occupant treatment and mean HCWs per day, were associated with increased HOB likelihood.
• Causal modeling confirmed the association between these non-patient features and HOB.

Conclusions:

• Non-individual patient features are valuable additions to machine learning models for HOB prediction.
• Integrating environmental and social interaction data can enhance the comprehensive analysis and prevention of HOB.
• These findings support a more holistic approach to understanding and mitigating hospital-acquired infections.