Supervised learning for infection risk inference using pathology data.

Area of Science:

• Biomedical Informatics
• Clinical Pathology
• Infectious Disease Epidemiology

Background:

• Antimicrobial resistance (AMR) is a growing threat to public health, exacerbated by the overuse of antimicrobials in hospitals.
• Clinical Decision Support Systems (CDSSs) can improve antimicrobial prescribing but may lead to unnecessary prescriptions without proper disease assessment.

Purpose of the Study:

• To develop a reliable method for infection risk inference using routinely available laboratory data.
• To support clinicians in making informed decisions regarding antimicrobial therapy initiation.

Main Methods:

• Combined six biochemical markers with microbiology susceptibility test outcomes from over 1.5 million daily patient profiles.
• Employed outlier detection, sampling techniques (SMOTE), and ten-fold cross-validation for robust model training and evaluation.
• Assessed model performance under conditions of missing data and class imbalance.

Main Results:

• Achieved high accuracy in infection risk inference with an area under the ROC curve of 0.80-0.83, sensitivity of 0.64-0.75, and specificity of 0.92-0.97.
• Standardization proved more effective than normalization, and SMOTE improved sensitivity.
• Models maintained performance with as few as four available biomarkers.

Conclusions:

• Selected biochemical markers provide sufficient information for confident infection risk inference, even with incomplete or imbalanced data.
• These findings can enhance CDSSs, enabling better antimicrobial stewardship and reducing unnecessary prescriptions.