OC-036 DEVELOPMENT OF NOVEL IMAGE-BASED DEEP LEARNING MODELS (DLMS) TO PREDICT OUTCOMES IN ABDOMINAL WALL RECONSTRUCTION (AWR) WITH IMBALANCED DATA SETS

Abstract

Abstract Aim The development of DLMs with imbalanced datasets represents a major challenge in the fields of artificial intelligence and surgery. The aim of this study was to develop and compare DLMs that predict mesh infection and pulmonary failure following AWR. Methods A prospectively maintained institutional database was used to identify AWR patients with preoperative CT scans. Standardized axial cuts of CT scans were rendered to DLMs. Conventional DLMs (CDLM) were developed two-class training system (i.e., learns negative and positive outcomes). CDLMs were compared to DLMs that were developed using a Generative Adversarial Network Anomaly (GANomaly) framework, which utilizes image augmentation and anomaly detection. The primary outcomes were receiver operating characteristic (ROC) values for predicting mesh infection and pulmonary failure. Results CT scans from 510 patients were utilized (10,004 images). Mesh infection and pulmonary failure occurred in 3.7% and 5.6% of patients, respectively. The CDLMs were less effective than GANomaly for predicting mesh infection (ROC 0.61 vs 0.73, p<0.01) and pulmonary failure (ROC 0.59 vs 0.70, p<0.01). Although the CDLMs had higher accuracies/specificities for predicting mesh infection (0.93 vs 0.78, p<0.01/0.96 vs 0.78, p<0.01) and pulmonary failure (0.88 vs 0.68, p<0.01/0.92 vs 0.67, p<0.01), they were substantially compromised by decreased model sensitivity (0.25 vs 0.68, p<0.01/ 0.27 vs 0.73, p<0.01). Conclusions Compared to CDLMs, GANomaly DLMs showed improved performance on imbalanced datasets, predominantly by increasing model sensitivity. Understanding patients who are at-risk for postoperative complications can improve risk stratification.