Predicting flap failure in head and neck surgery: Data Augmentation and Resampling for Data Imbalance

Abstract

Abstract Objective The area under receiver operating characteristic curve (AUROC) is hampered by imbalanced data using artificial intelligent models for predicting free flap failure. Thus, we propose a new data preprocessing method with Gaussian Noise and Up-Sampling to increase the AUROC score. Study design: Case-control study Setting: Data were obtained from patients with head and neck cancer who underwent free flap reconstruction at Changhua Christian Hospital in Taiwan between May 2019 and June 2020. Methods The collected dataset was initially divided into training and validation sets. Subsequently, a data augmentation technique was employed on the training dataset to generate additional training data, thereby addressing the issue of imbalanced sample sizes between successful and unsuccessful outcomes. The rebalanced training data are then utilized to optimize the parameters of diverse machine learning frameworks, including logistic regression(LR), random forest (RF), support vector machine (SVM), ensemble models, and multi-layer perceptron neural network (MLP). Result The AUROC values derived from the original data were respectively 0.50 for the SVM and 0.53, RF, and augmented data were significantly improved to 0.72 for SVM and LR, 0.58, and 0.57, RF. When the features “Age,’ ’Preoperative Hb level,’ and” total blood loss’ were removed, the dropped AUROC value was 0.3, indicating their highest importance. Conclusion The data augmentation method adopted in this study resolves the data imbalance problem and improves the efficacy of most machine learning models used to predict free flap failure in head and neck surgery.