American Association for Thoracic Surgery Quality Gateway Procedural Outcomes Models for General Thoracic Surgery...

To develop quality assurance models for adverse procedural events after lung resection and esophagectomy for cancer.

Using procedural outcomes after 16,430 lung resections and 3,068 esophagectomies for cancer at 9 hospitals in 7 hospital systems, we developed 5 individual and 2 composite outcome models from preoperative information. Random forest quantile classification for imbalanced data with model-free variable priority screening identified predictors. Out-of-bag prediction provided regularized estimates and internal validation. Hospital-holdout analysis assessed institutional benchmark stability.

Of 71 lung resection and 63 esophagectomy preoperative variables, 35 and 29, respectively, were identified as important predictors across the 7 outcome models. These included demographics, advanced cancer, organ dysfunction, higher acuity, and planned surgical approach and resection extent. Out-of-bag calibration covered a modest range of probabilities. Out-of-bag geometric mean ranged from 0.588 to 0.713 for lung resection and 0.517 to 0.649 for esophagectomy. Continuous risk factors, particularly pulmonary function, provided high-importance information content, with complex dependency and interaction structure among predictors. Risk-factor magnitude differed among hospitals, as did their risk-adjusted outcomes according to patient risk as determined by counterfactual inference within a framework of virtual twins. For lung resection operative mortality, hospital-specific contrasts were stable when each focal hospital was excluded from the aggregate benchmark.

Using contemporary machine learning, initial models for adverse procedural events after lung resection and esophagectomy for cancer have been developed. They provide surgeons, their patients, and hospital and hospital systems tools for patient risk assessment and for informing quality improvement initiatives in general thoracic surgery.