2168 Machine Learning for Cost Prediction in Cervical Fusion: A Novel Application of Time-Driven Activity-Based Costing

Le résumé fourni par la source

INTRODUCTION: The rising costs of spine surgery necessitate improved financial planning and resource allocation. Although studies have used time-driven activity-based costing (TDABC) to quantify the true costs of care, its integration with machine learning for predictive modeling remains unexplored in cervical fusion procedures. METHODS: Data from 722 cervical fusion procedures (ACDF or PCDF) performed at our institution between 2017-2022 were analyzed. Variables included number of levels, surgeon, preoperative NDI, procedure indication (myelopathy, radiculopathy, etc.), surgical approach (ACDF or PCDF), patient demographics, and comorbidities. Linear regression, Random Forest, and Gradient Boosting were used for total cost prediction. Logistic regression, Random Forest, Gradient Boosting, Decision Tree, and Neural Network models were employed for high-cost case identification (>93rd percentile, 1.5 SD above mean). Model performance was assessed using R2 and AUC. RESULTS: For overall cost prediction, linear regression achieved the highest R2 (0.70), while Random Forest performed best for typical cases (R2 = 0.80 for costs below 93rd percentile). However, all models performed poorly for high-cost outliers (R2 = 0). For high-cost identification, Random Forest demonstrated superior performance (AUC = 0.874), followed by Neural Network (AUC = 0.853). Key predictors of high costs identification included number of surgical levels, surgeon, procedure type, and preoperative NDI. CONCLUSIONS: Machine learning models integrated with TDABC can effectively predict costs for typical cervical fusion procedures, but struggles with high-cost outliers. Nonetheless, machine learning models can accurately predict when surgeries will become a high-cost surgery. This approach enables preoperative identification of potential cost drivers and supports financial planning for most cases. The findings highlight the potential of predictive analytics in enhancing cost management and resource allocation in spine surgery while recognizing the challenge of forecasting unpredictable intraoperative and patient-specific factors that drive extreme costs.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.