Prediction of postoperative infections by strategic data imputation and explainable machine learning

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Surgical Outcomes Research

Background:

• Healthcare-associated infections (HAIs) are a major cause of patient morbidity and mortality.
• Early detection of postoperative infections is critical for effective patient management.
• Traditional predictive models for surgical outcomes primarily rely on preoperative data.

Purpose of the Study:

• To evaluate the efficacy of integrating postoperative laboratory values and their kinetics for predicting bacterial infections.
• To develop and assess procedure-agnostic prediction models incorporating dynamic laboratory data.
• To compare the performance of dynamic models against static or preoperative-only models.

Main Methods:

• Analysis of 91,794 surgical cases from electronic health records (EHR).
• Development of machine learning models using preoperative, intraoperative, and postoperative variables.
• Incorporation of static and kinetic properties of laboratory values, with strategic imputation for missing data.

Main Results:

• Machine learning models integrating laboratory value kinetics achieved recall, precision, and ROC AUC of 0.71, 0.69, and 0.83, respectively, by postoperative day 2.
• Infection detection using dynamic models outperformed clinician-based decision-making.
• Identification of novel predictive combinations from hepatic, renal, and bone marrow function markers.

Conclusions:

• Dynamic modeling of postoperative laboratory values enhances the timeliness and accuracy of infection detection.
• Explainable machine learning aids clinical interpretation and highlights multi-organ system involvement in infection risk.
• The developed workflow is surgery-independent, generalizable, and has potential to optimize patient outcomes and resource utilization.