SADA: enhanced framework for predicting surgical risk via adaptive data augmentation and high-risk identification

Abstract

Accurate surgical risk prediction has a positive effect on clinical resource planning, preparation of contingency plans, and reduction of surgical risk for patients. The American Society of Anesthesiologists Physical Status (ASA-PS) is the most widely used method for predicting surgical risk. Consequently, various studies using machine learning to predict ASA risk have arisen. However, previous studies have failed to address the adverse effects of unbalanced surgical data on prediction performance and have not given enough consideration to the rare but valuable subset of high-risk surgery. In order to better estimate ASA surgical risk, this paper proposes a Self-Adaptive Data Augmentation Framework for Enhanced Risk Prediction (SADA), which attempts to efficiently balance data while concentrating on high-risk operations. SADA employs a two-stage prediction framework: first, in the first stage, the AdaptiveRisk AutoEncoder is proposed for balancing surgical data and identifying high-risk procedures. Specifically, the data in the encoder part undergoes an adaptive balancing process, and a classification layer is added to the decoder part so that the model concentrates on infrequent high-risk surgeries; second, a machine learning model is employed in the second stage to predict the ASA grade of low-risk surgeries in order to give doctors detailed and precise ASA prediction information. The experimental findings demonstrate the superiority of SADA in surgical risk prediction compared to models often employed in surgical risk prediction.