Bundled Care for Hip Fractures: A Machine-Learning Approach to an Untenable Patient-Specific Payment Model

With the transition to a value-based model of care delivery, bundled payment models have been implemented with demonstrated success in elective lower extremity joint arthroplasty. Yet, hip fracture outcomes are dependent on patient-level factors that may not be optimized preoperatively due to acuity of care. The objectives of this study are to (1) develop a supervised naive Bayes machine-learning algorithm using preoperative patient data to predict length of stay and cost after hip fracture and (2) propose a patient-specific payment model to project reimbursements based on patient comorbidities.

Using the New York Statewide Planning and Research Cooperative System database, we studied 98,562 Medicare patients who underwent operative management for hip fracture from 2009 to 2016. A naive Bayes machine-learning model was built using age, sex, ethnicity, race, type of admission, risk of mortality, and severity of illness as predictive inputs.

Accuracy was demonstrated at 76.5% and 79.0% for length of stay and cost, respectively. Performance was 88% for length of stay and 89% for cost. Model error analysis showed increasing model error with increasing risk of mortality, which thus increased the risk-adjusted payment for each risk of mortality.

Our naive Bayes machine-learning algorithm provided excellent accuracy and responsiveness in the prediction of length of stay and cost of an episode of care for hip fracture using preoperative variables. This model demonstrates that the cost of delivery of hip fracture care is dependent on largely nonmodifiable patient-specific factors, likely making bundled care an implausible payment model for hip fractures.