OBJECTIVES

Explore multi-level predictors of 90-day post-acute costs after elective total knee arthroplasty(TKA) and evaluate performance of predictive models for classification of high-cost patients.

METHODS

Using the Medicare Standard Analytic File(SAF), 90-day post-acute standardized costs were quantified for patients who underwent elective TKA between October 2015 and June 2017. Patient demographics, comorbidities, prior healthcare utilization, hospital characteristics, surgeon variables, and census data were included in Generalized Estimating Equation(GEE) logit regression and random forest(RF) models to identify patients in the top decile of post-acute costs. LASSO regression was used with the GEE model, which considered clustering of patients within hospitals. A 70/30 random split was used for training models and testing performances.

RESULTS

302,440 patients were included, the majority were female(63%) with a mean age of 74. Mean(SD) post-acute 90-day cost was $5,419($7,610)-the top decile was $12,778. The top 5 predictors in GEE were: provider geography, neurological disorders, psychosis, eligibility for both Medicare and Medicaid, and paralysis. The top 5 in RF were: age, proportion patients discharged home, cumulative physician TKA experience, provider annual TKA volume, and hospital socioeconomic profile. Prediction performance, based on area under the receiver operating characteristic(ROC) curve, was 0.76 for RF compared to 0.74 for GEE.

CONCLUSIONS

302,440 patients were included, the majority were female(63%) with a mean age of 74. Mean(SD) post-acute 90-day cost was $5,419($7,610)-the top decile was $12,778. The top 5 predictors in GEE were: provider geography, neurological disorders, psychosis, eligibility for both Medicare and Medicaid, and paralysis. The top 5 in RF were: age, proportion patients discharged home, cumulative physician TKA experience, provider annual TKA volume, and hospital socioeconomic profile. Prediction performance, based on area under the receiver operating characteristic(ROC) curve, was 0.76 for RF compared to 0.74 for GEE.