Development and Validation of Image-Based Deep Learning Models to Predict Surgical Complexity and Complications in Abdominal Wall Reconstruction

Abstract

Image-based deep learning models (DLMs) have been used in other disciplines, but this method has yet to be used to predict surgical outcomes. To apply image-based deep learning to predict complexity, defined as need for component separation, and pulmonary and wound complications after abdominal wall reconstruction (AWR). This quality improvement study was performed at an 874-bed hospital and tertiary hernia referral center from September 2019 to January 2020. A prospective database was queried for patients with ventral hernias who underwent open AWR by experienced surgeons and had preoperative computed tomography images containing the entire hernia defect. An 8-layer convolutional neural network was generated to analyze image characteristics. Images were batched into training (approximately 80%) or test sets (approximately 20%) to analyze model output. Test sets were blinded from the convolutional neural network until training was completed. For the surgical complexity model, a separate validation set of computed tomography images was evaluated by a blinded panel of 6 expert AWR surgeons and the surgical complexity DLM. Analysis started February 2020. Image-based DLM. The primary outcome was model performance as measured by area under the curve in the receiver operating curve (ROC) calculated for each model; accuracy with accompanying sensitivity and specificity were also calculated. Measures were DLM prediction of surgical complexity using need for component separation techniques as a surrogate and prediction of postoperative surgical site infection and pulmonary failure. The DLM for predicting surgical complexity was compared against the prediction of 6 expert AWR surgeons. A total of 369 patients and 9303 computed tomography images were used. The mean (SD) age of patients was 57.9 (12.6) years, 232 (62.9%) were female, and 323 (87.5%) were White. The surgical complexity DLM performed well (ROC = 0.744; P < .001) and, when compared with surgeon prediction on the validation set, performed better with an accuracy of 81.3% compared with 65.0% (P < .001). Surgical site infection was predicted successfully with an ROC of 0.898 (P < .001). However, the DLM for predicting pulmonary failure was less effective with an ROC of 0.545 (P = .03). Image-based DLM using routine, preoperative computed tomography images was successful in predicting surgical complexity and more accurate than expert surgeon judgment. An additional DLM accurately predicted the development of surgical site infection.