Estimation of Risk-Adjusted Postoperative Infection Outcomes Using Interpretable Machine Learning and Electronic Health Record Data.

Background: This study compared risk-adjusted postoperative infection outcomes estimated by statistical models applied to electronic health record (EHR) data ("automated") to gold-standard manual chart review outcomes estimated by the American College of Surgeons National Surgical Quality Improvement Program (ACS-NSQIP).

Materials and methods: A cohort of adult patients who had operations in nine surgical specialties at five large hospitals within one healthcare system between 2013-2019 were included. 307,335 patients underwent 441,047 unique operations. Records from 30,603 patients were linked to the local ACS-NSQIP database (97% linkage). Previously published models for estimating preoperative risk and occurrence of postoperative infections were used to estimate observed to expected event ratios (O/E) for surgical site infections, urinary tract infections, sepsis/septic shock, and pneumonia.

Results: Risk-adjusted infection outcomes expressed as O/E ratios were similar when comparing EHR automated methods to manual chart review across five hospitals and four infection types. The Pearson correlation coefficient of the hospital O/E ratios was 0.77, mean absolute difference was 0.13, and 100% of the confidence intervals were overlapping. The correlations and mean absolute differences for individual infection types improved as incidence rates increased.

Discussion: Parsimonious statistical models applied to EHR data can be used to accurately estimate hospital risk-adjusted postoperative infection outcomes for all operations.

Conclusions: These models could be used to augment postoperative infection surveillance for hospital quality monitoring.