The use of machine learning for predicating complications of free flap head and neck reconstruction.

Abstract

e18633 Background: The use of machine learning (ML) in plastic and reconstructive surgery has increased over the last decade. However, its use to predict surgical outcomes in head and neck reconstruction has not been well studied. The goal of this study is to assess the performance of ML algorithms trained to predict surgical outcomes of head and neck free flap reconstruction. Methods: Our study cohort included routinely collected data from 4000 patients who underwent free flaps for reconstruction of head and neck defects between January 2005 and December 2018. We developed and tested nine supervised ML algorithms to predict three outcomes of a.) any complication, b.) any major recipient-site complication, and c.) total flap loss. Results: In our sample, 33.7% of patients experienced any complication, 26.5% experienced a major complication at the recipient site and 1.7% experienced total flap loss. The k-nearest neighbors algorithm demonstrated the best overall performance for predicting any complication (AUROC = .61, sensitivity = .60). Regularized regression had the best performance for predicting major recipient site complication (AUROC=.68, sensitivity = .66), and decision trees were the best predictors of total flap loss (AUROC = .66, sensitivity = .50). Conclusions: We demonstrated that ML models trained using routinely collected data can make clinical useful predictions about who will experience complications. Our models correctly identified between half and two-thirds of patients who experienced post-surgical complications including total flap loss. These models can be applied to readily available clinical and perioperative data to facilitate decision making. Further performance improvements are likely possible with the inclusion of additional variables related to patient health and behavior.