Predicting rare outcomes in abdominal wall reconstruction using image-based deep learning models.

Area of Science:

• Artificial intelligence in surgery
• Medical imaging analysis
• Deep learning for predictive modeling

Background:

• Imbalanced datasets pose challenges for artificial intelligence (AI) and surgical deep learning models.
• Predicting rare postoperative complications after abdominal wall reconstruction is crucial but difficult.

Purpose of the Study:

• To develop and compare deep learning models for predicting rare postoperative complications in abdominal wall reconstruction.
• To evaluate the efficacy of generative adversarial network (GAN) anomaly detection against conventional deep learning approaches.

Main Methods:

• Utilized a prospectively maintained institutional database of abdominal wall reconstruction patients with preoperative computed tomography (CT) scans.
• Developed conventional 8-layer convolutional neural network (CNN) models and GAN anomaly framework models for prediction.
• Compared model performance using receiver operating characteristic (ROC) values for mesh infection and pulmonary failure prediction.

Main Results:

• GAN anomaly models outperformed conventional models in predicting mesh infection (ROC 0.73 vs 0.61) and pulmonary failure (ROC 0.70 vs 0.59).
• Conventional models showed higher accuracy/specificity but significantly lower sensitivity (0.25-0.27) compared to GAN anomaly models (0.68-0.73).
• GAN anomaly models demonstrated improved performance on imbalanced data, primarily through enhanced model sensitivity.

Conclusions:

• Generative adversarial network anomaly deep learning models offer superior performance for predicting rare complications in imbalanced surgical datasets.
• Improved prediction of rare complications can enhance patient risk stratification, resource allocation, and informed consent processes.
• This AI-driven approach holds potential for improving outcomes in abdominal wall reconstruction surgery.