The impact of complication-sensitive risk models on hospital benchmarking for failure to rescue

Abstract

Background

Failure to rescue has been increasingly used as a surgical quality metric, although implementation with complication-agnostic risk models may disproportionately penalize centers that care for high-risk patients. We used a nationally representative database to assess the impact of complication-sensitive risk models on hospital benchmarking for failure to rescue.

Methods

All adults undergoing elective coronary artery bypass grafting, aortic/mitral valve replacement, or esophageal/pancreatic/large bowel resection were identified within the 2019 Nationwide Readmissions Database. Two hierarchical logistic regressions (model 1: complication-agnostic; model 2: complication-sensitive) were developed to evaluate risk-adjusted rates of failure to rescue at each center. Patient characteristics (demographics, comorbidities) were incorporated as fixed effects in both models. Model 2 also included adjustment for the occurrence and identity of each complication. Hospitals were subsequently grouped into quintiles of failure to rescue using each model.

Results

Approximately 296,907 patients at 1,034 hospitals met inclusion criteria. Overall mortality, complication, and failure to rescue rates were 1.1%, 4.8%, and 17.8%, respectively. Centers in the highest quintile of failure to rescue for model 1 more frequently managed patients who developed cardiac arrest (0.9 vs 0.7%, P = .003) or acute kidney injury requiring dialysis (0.6 vs 0.4%, P = .017). In contrast, the rates of all complications except sepsis (2.7 vs 2.3%, P = .035) were comparable between centers in the top quintile and others, when using model 2. Overall, ∼30% of hospitals were reclassified into different quintiles with the complication-sensitive model.

Conclusion

This study suggests that complication-agnostic models disproportionately penalize centers caring for patients who develop severe complications, which can be mitigated with complication-sensitive models.