Optimal Computed Tomography-based Biomarkers For Prediction Of Incisional Hernia Formation
Ankoor A. Talwar, MBA, Phoebe B. McAuliffe, BS, Abhishek A. Desai, MD, Robyn B. Broach, PhD, Jesse Y. Hsu, PhD, Tiange Liu, PhD, Jayaram K. Udupa, PhD, Yubing Tong, PhD, Drew A. Torigan, PhD, Joseph M. Serletti, MD, John P. Fischer, MD MPH.
University of Pennsylvania, Philadelphia, PA, USA.
PURPOSE: Incisional hernia (IH) is a pervasive surgical disease. Computer vision and machine learning (ML) were used to derive computed tomography (CT)-based features predictive of IH.
METHODS: Patients who underwent colorectal surgery between 2005-2017 were identified (n=14,345). Patients who developed IH were matched with those who did not (n=212). Preoperative abdominopelvic CT scans were segmented to derive linear, volumetric, intensity-based, and textural measurements (Figure 1). Optimal biomarkers (OBMs) were derived and used to test ML classifiers (SVMs, Random Forests, Ensemble Boosting) for IH prediction.
RESULTS: 279 features were extracted from preoperative CTs. The most predictive combination of OBMs was: 1) abdominopelvic visceral adipose tissue volume (VAT); 2) abdominopelvic intra-abdominal musculature volume; and 3) pelvic VAT to outer abdominal musculature volume (OAM) ratio (Figure 2). ML models using these OBMs were tested, Ensemble Boosting outperformed other models across all metrics (Figure 3).
CONCLUSION: These OBMs suggest intra-abdominal volume/pressure is the most salient pathophysiologic mechanism for IH formation. ML models using OBMs are highly predictive of IH. Image analysis is a powerful tool in surgical risk prediction.
Back to 2023 Abstracts