Classifying Infection Risk Following Pediatric Cardiac Surgery

Area of Science:

• Pediatric cardiac surgery
• Infectious disease epidemiology
• Health informatics

Background:

• Postoperative infections are a significant complication in pediatric cardiac surgery, leading to increased morbidity and healthcare costs.
• Early identification of high-risk patients could enable targeted preventive strategies to mitigate infection risk.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting postoperative infections in pediatric cardiac surgery.
• To compare the performance of models trained on validated versus unvalidated outcome data.

Main Methods:

• A cohort of 2080 pediatric cardiac surgery cases was created using structured health data from a single center.
• Outcomes (sepsis, bacteremia, necrotizing enterocolitis, composite outcome) were initially assigned using billing codes and subsequently validated manually.
• Logistic regression and machine learning methods were employed to train predictive models.

Main Results:

• Manual validation indicated low accuracy of diagnosis codes for classifying postoperative infections.
• Despite discrepancies in outcome assignments between validated and unvalidated data, predictive model performance was comparable.
• Machine learning models showed similar efficacy regardless of outcome data validation status.

Conclusions:

• Billing codes alone are insufficient for accurately identifying postoperative infections in pediatric cardiac surgery.
• Machine learning models for infection risk prediction can achieve similar performance using either unvalidated or validated outcome data, simplifying data acquisition.
• Further research is needed to refine predictive models and improve the accuracy of outcome ascertainment in electronic health records.