Division of Gastrointestinal and Minimally Invasive Surgery, Department of Surgery, Carolinas Medical Center, Charlotte, NC.
Graduate School, Department of Statistics, Columbia University, New York, NY.
Surgery. 2023 Mar;173(3):748-755. doi: 10.1016/j.surg.2022.06.048. Epub 2022 Oct 11.
Deep learning models with imbalanced data sets are a challenge in the fields of artificial intelligence and surgery. The aim of this study was to develop and compare deep learning models that predict rare but devastating postoperative complications after abdominal wall reconstruction.
A prospectively maintained institutional database was used to identify abdominal wall reconstruction patients with preoperative computed tomography scans. Conventional deep learning models were developed using an 8-layer convolutional neural network and a 2-class training system (ie, learns negative and positive outcomes). Conventional deep learning models were compared to deep learning models that were developed using a generative adversarial network anomaly framework, which uses image augmentation and anomaly detection. The primary outcomes were receiver operating characteristic values for predicting mesh infection and pulmonary failure.
Computed tomography scans from 510 patients were used with a total of 10,004 images. Mesh infection and pulmonary failure occurred in 3.7% and 5.6% of patients, respectively. The conventional deep learning models were less effective than generative adversarial network anomaly for predicting mesh infection (receiver operating characteristic 0.61 vs 0.73, P < .01) and pulmonary failure (receiver operating characteristic 0.59 vs 0.70, P < .01). Although the conventional deep learning models had higher accuracies/specificities for predicting mesh infection (0.93 vs 0.78, P < .01/.96 vs .78, P < .01) and pulmonary failure (0.88 vs 0.68, P < .01/.92 vs .67, P < .01), they were substantially compromised by decreased model sensitivity (0.25 vs 0.68, P < .01/.27 vs .73, P < .01).
Compared to conventional deep learning models, generative adversarial network anomaly deep learning models showed improved performance on imbalanced data sets, predominantly by increasing model sensitivity. Understanding patients who are at risk for rare but devastating postoperative complications can improve risk stratification, resource utilization, and the consent process.
在人工智能和外科领域，数据不均衡的深度学习模型是一个挑战。本研究旨在开发和比较预测腹壁重建术后罕见但严重并发症的深度学习模型。
使用前瞻性维护的机构数据库来识别接受腹壁重建且术前进行计算机断层扫描的患者。使用 8 层卷积神经网络和 2 类训练系统（即，学习阴性和阳性结果）开发传统的深度学习模型。将传统的深度学习模型与使用生成对抗网络异常框架开发的深度学习模型进行比较，该框架使用图像增强和异常检测。主要结果是预测网感染和肺衰竭的受试者工作特征值。
共使用 510 例患者的计算机断层扫描，共 10004 张图像。网感染和肺衰竭的发生率分别为 3.7%和 5.6%。与生成对抗网络异常相比，传统的深度学习模型预测网感染（受试者工作特征 0.61 与 0.73，P <.01）和肺衰竭（受试者工作特征 0.59 与 0.70，P <.01）的效果较差。尽管传统的深度学习模型预测网感染（0.93 与 0.78，P <.01/.96 与.78，P <.01）和肺衰竭（0.88 与 0.68，P <.01/.92 与.67，P <.01）的准确性/特异性更高，但模型敏感性降低（0.25 与 0.68，P <.01/.27 与.73，P <.01）。
与传统的深度学习模型相比，生成对抗网络异常深度学习模型在不均衡数据集上的表现更好，主要通过提高模型敏感性。了解有罕见但严重术后并发症风险的患者可以改善风险分层、资源利用和知情同意过程。
