CSA Home
Central Surgical Association

Back to 2022 Abstracts


Predicting Rare Outcomes In Abdominal Wall Reconstruction (AWR) Using Novel Image-based Artificial Intelligence (AI) And Deep Learning Models (DLM)
*Sullivan A Ayuso1, *Sharbel A Elhage1, *Yizi Zhang2, Keith S Gersin1, *Vedra A Augenstein1, *Paul D Colavita1, *B Todd Heniford1
1Carolinas Medical Center, Charlotte, NC;2Columbia University, New York, NY

OBJECTIVE(S): AI and DLMs have only recently been applied to surgical outcome prediction. Previously, AI was predominantly used for disease diagnosis. Developing DLMs using imbalanced datasets, in which there are none or few examples of the minority class, is a major challenges in AI prediction. The aim of this study was to develop and compare novel DLMs that predict rare but significant postoperative complications for patient undergoing AWR.
METHODS: A prospective institutional database was used to identify AWR patients with preoperative CT scans. Conventional DLMs (CDLM) were developed using an 8-layer convolutional neural network and a two-class training system (i.e., learns negative and positive outcomes). CDLMs were compared to DLMs that were developed with Generative Adversarial Network Anomaly Detection (GANomaly) models, which learn negative outcomes and then recognize positive outcomes as anomalies. The primary outcomes were area under the curve (AUC) for predicting mesh infection and pulmonary insufficiency/failure requiring ICU care.
RESULTS: Out of the 2,698 open AWR patients in our single-institution database, CT scans from 511 patients were utilized (total of 10,004 images). Mesh infection occurred in 3.7% of patients (462 images), and pulmonary failure occurred in 5.6% of patients (720 images). For both mesh infection and pulmonary failure, there were 408 patients (7,831 images) in the training set and 103 patients in the test set. The CDLMs were less effective than the GANomaly DLMS for predicting mesh infection (AUC 0.61vs0.73; p<0.01) and pulmonary failure (AUC 0.59vs0.70; p<0.01). Although the CDLMs had higher accuracies/specificities for predicting mesh infection (0.93vs0.78; p<0.01 and 0.96vs0.78; p<0.01) and pulmonary failure (0.88vs0.68; p<0.01 and 0.92vs0.67; p<0.01), they were compromised by a significant decrease in model sensitivity (0.25vs0.68; p<0.01 and 0.27vs0.73; 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 potentially influence perioperative decision making, informed consent, resource utilization, and be applied to other areas in medicine to predict rare outcomes.


Back to 2022 Abstracts