Comparison of Machine Learning-Based Reporting with Surgeon Reporting of Postoperative Complication in Cardiothoracic Morbidity and Mortality Conferences.

BACKGROUND

Traditional morbidity and mortality (M&M) conferences incompletely capture postoperative complications, potentially limiting quality improvement efforts. We developed and validated the Automated Surveillance of Postoperative Infectious and Non-Infectious Complications (ASPIN), a machine-learning system that estimates postoperative complication rates from electronic health record data, and compared ASPIN estimates with surgeon-reported complications from cardiothoracic M&M conferences.

STUDY

DESIGN

Cardiothoracic M&M reports from January 1, 2022, through December 31, 2024, were manually reviewed to identify surgeon-reported complications. Reports were matched to electronic health records by medical record number, operation date, and primary surgeon. Patient characteristics were compared using Wilcoxon rank-sum and chi-squared tests, and surgeon-reported versus ASPIN-estimated complication rates were compared using paired t-tests.

RESULTS

Among 4,522 cardiac and thoracic operations, surgeons reported ≥1 complication in 834 cases (18.4%), whereas ASPIN estimated an overall complication rate of 32.8%. ASPIN estimated higher rates for most complications, with the largest differences observed for bleeding requiring transfusion (17.68% vs 1.15%), sepsis (9.57% vs 0.29%), surgical site infection (9.02% vs 0.46%), and pneumonia (7.40% vs 0.95%) (all p<0.001). Lower ASPIN estimates were observed for cardiac complications (4.33% vs 5.06%; p=0.042), renal complications (0.75% vs 1.64%; p<0.001), and readmission (0.96% vs 1.81%; p<0.001).

CONCLUSIONS

Compared with traditional M&M reporting, ASPIN identified substantially higher postoperative morbidity, particularly for bleeding, sepsis, surgical site infection, and pneumonia. Automated surveillance may complement conventional M&M processes by improving complication detection, case selection, and postoperative quality monitoring.