Background: Risk-prediction indices are one category of the many tools implemented to guide efforts to decrease readmissions. However, using fied models to predict a complex process can prove challenging. In addition, no risk-prediction index has been developed for patients undergoing colorectal surgery. Therefore, we evaluated the performance of a widely utilized simplified index developed at the hospital level - LACE (length of stay, acute admission, Charlson comorbidity index score, and emergency department visits) and developed and evaluated a novel index in predicting readmissions in this patient population.
Methods: Using a retrospective split-sample cohort, patients discharged after colorectal surgery were identified within the inpatient databases of the Healthcare Cost and Utilization Project for the states of New York, California, and Florida (2006-2014). The primary outcome was death or readmission within 30 days after discharge. Multivariable logistic regression models incorporated patient comorbidities, postoperative complications, and hospitalization details, and were evaluated using the C statistic.
Results: A total of 440,742 patients met eligibility criteria. The rate of death or readmission within 30 days after discharge was 14.0% (n = 61,757). When applied to surgical patients, the LACE index demonstrated a poor model fit (C = 0.631). The model fit improved significantly-but remained poor (C = 0.654; P < .001)-with the addition of the following variables, which are known to be associated with readmission after colorectal surgery: age, indication for surgery, and creation of a new ostomy. A novel, simplified model also yielded a poor model fit (C = 0.660).
Conclusion: Postdischarge death or readmission after colorectal surgery is not accurately modeled using existing, modified, or novel simplified risk prediction models. Payers and providers must ensure that quality improvement efforts applying simplified models to complex processes, such as readmissions following colorectal surgery, may not be appropriate, and that models reflect the relevant patient population.
Published by Elsevier Inc.