K. R. Chhabra1,3,4, U. Nuliyalu4, J. B. Dimick2,3,4, H. Nathan2,3,4 1Brigham And Women’s Hospital,Department Of Surgery,Boston, MA, USA 2University Of Michigan,Department Of Surgery,Ann Arbor, MI, USA 3University Of Michigan,IHPI Clinician Scholars Program,Ann Arbor, MI, USA 4University Of Michigan,Center For Healthcare Outcomes And Policy,Ann Arbor, MI, USA
Introduction:
Surgery accounts for almost half of inpatient spending, much of which is concentrated in a subset of high cost patients. A method of prospectively identifying high cost patients, i.e. “hot spotting,” may help manage population health spending, but we lack an optimal way to predict which patients will have high-cost surgical episodes.
Methods:
Using 100% Medicare claims data, we identified patients aged 65-99 undergoing elective inpatient surgery (CABG, colectomy, total hip/knee replacement) in 2014. We calculated price-standardized Medicare payments for the surgical episode from admission through 30 days after discharge (episode payments). Based on predictor variables from 2013, e.g. Elixhauser comorbidities, hierarchical condition categories, Medicare’s Chronic Conditions Warehouse (CCW), and total spending, we constructed models to predict the costs of surgical episodes in 2014. We used general linear mixed models incorporating hospital random effects and adjusting for age, sex, and race, testing fit with R2 and kappa statistics (κ) using quintiles of spending.
Results:
A cost prediction model based on CCW score performed well in predicting payment variation for all procedures (R2 0.16-0.22, κ 0.13-0.15; all P<0.001). Other models also had statistically significant R2 and κ but had inferior predictive performance to CCW. The costliest quintile of patients as predicted by the model captured 40-50% of the patients in each procedure’s actual costliest quintile. For example, in CABG, 48% of the costliest quintile was predicted by the model’s costliest quintile. A greater share of the costliest quintile was identified when the prediction threshold was lowered; e.g. in CABG 73% of the actual costliest quintile was identified by combining the model’s 2 top quintiles of predicted cost.
Conclusion:
Expensive surgical patients can be prospectively identified using readily available data on patients’ chronic conditions. The sensitivity of the cost prediction model can be tailored as desired. For instance, if attempting to identify as many potentially expensive patients as possible, one may lower the threshold for detection by combining quintiles of predicted cost.
