Can a machine learning model accurately predict patient resource utilization following lumbar spinal fusion?

Area of Science:

• Spine Surgery
• Health Economics
• Machine Learning

Background:

• Value-based healthcare and bundled payments are increasingly adopted by the Centers for Medicare and Medicaid Services.
• Patient variability in spine surgery challenges the viability of uniform bundled payment models.
• Machine learning offers potential for predicting patient-specific outcomes in spine surgery to inform payment schemes.

Purpose of the Study:

• To determine if a Naïve Bayes machine-learning model can predict inpatient payments, length of stay (LOS), and discharge disposition for dorsal and lumbar fusion.
• To assess the utility of such a model in developing risk-stratified bundled payment schemes.

Main Methods:

• A Naïve Bayes machine-learning model was developed using an administrative database.
• The model included patients undergoing dorsal and lumbar fusion for nondeformity indications (2009-2016).
• Predictive inputs included patient demographics, admission type, and clinical severity codes; outcomes were LOS, discharge disposition, and total inpatient payments.

Main Results:

• The Naïve Bayes model demonstrated high reliability (AUCs 0.880-0.941) in predicting cost, LOS, and discharge disposition.
• Patients with higher risk of mortality or severity of illness had greater inpatient payments.
• Individualized, risk-based payment models are warranted due to the wide range of expected payments based on preoperative comorbidities.

Conclusions:

• A Naïve Bayes model reliably predicts resource utilization outcomes for spine fusion surgery.
• Significant cost differences exist across risk strata (APR risk of mortality and severity of illness).
• Risk-adjusted payment plans mitigate financial risk for institutions treating high-comorbidity patients, unlike flat bundled payments.