Interpretable machine learning accurately reclassifies lobectomy surgical approaches by cost

Abstract

BackgroundThe volume of robotic lung resection continues to increase despite its higher costs and unproven superiority to video-assisted thoracoscopic surgery. We evaluated whether machine learning can accurately identify factors influencing cost and reclassify high-cost operative approaches into lower-cost alternatives. MethodsThe Florida Agency for Healthcare Administration and Centers for Medicare and Medicaid Services Hospital and Physician Compare datasets were queried for patients undergoing open, video-assisted thoracoscopic surgery and robotic lobectomy. K-means cluster analysis was used to identify robotic clusters based on total cost. Predictive models were built using artificial neural networks, Support Vector Machines, Classification and Regression Trees, and Gradient Boosted Machines algorithms. Models were applied to the high-volume robotic group to determine patients whose cost cluster changed if undergoing a video-assisted thoracoscopic surgery approach. A local interpretable model-agnostic explanation approach personalized cost per patient. ResultsOf the 6,618 cases included in the analysis, we identified 4 cost clusters. Application of artificial neural networks to the robotic subgroup identified 1,642 (65%) cases with no re-assignment of cost cluster, 583 (23%) with reduced costs, and 300 (12%) with increased costs if they had undergone video-assisted thoracoscopic surgery approach. The 5 overall highest cost predictors were patient admission from the clinic, diagnosis of metastatic cancer, presence of cancer, urgent hospital admission, and dementia. ConclusionK-means cluster analysis and machine learning identify a patient population that may undergo video-assisted thoracoscopic surgery or robotic lobectomy without a significant difference in total cost. Local interpretable model-agnostic explanation identifies individual patient factors contributing to cost. Application of this modeling may reliably stratify high-cost patients into lower-cost approaches and provide a rationale for reducing expenditure.