Eric S Wise, MD, Keaton Joppru, BS, Adam Sheka, MD, Keith Wirth, MD, Stuart Amateau, MD, Sayeed Ikramuddin, MD, Daniel B Leslie, MD. University Of Minnesota
Introduction: Multiple patient factors may convey increased risk of complications and death after bariatric surgery. Assessing the likelihood of short-term morbidity is useful for both the bariatric surgeon and patient. Artificial neural networks are computational models that use pattern recognition to improve outcome predictions, and can continually be refined to reflect new data. These advanced models, coupled with a robust national bariatric database, can be useful in answering critical questions about those factors governing postoperative morbidity. This study aims to implement an artificial neural network model to optimize prediction of 30-day major morbidity and mortality after laparoscopic sleeve gastrectomy (LSG), using only simple, readily known preoperative patient factors.
Methods and Procedures: 101,721 patients who underwent LSG were considered for analysis from the 2016 Metabolic and Bariatric Surgery Accreditation and Quality Improvement Program national database. Select patient factors were selected a priori that were deemed simple, pertinent and easily obtainable, and their association with 30-day major complications (readmission, reintervention, reoperation or death) was assessed. Those factors with a significant association on both univariate and multivariate nominal logistic regression analysis were incorporated into a backpropagation neural network with 3 nodes each assigned a training value of 0.333, with k-fold internal validation. Logistic regression and ANN models were compared using area under receiver-operating characteristic curves (AUROC). A criterion of P≤0.05 was used to denote statistical significance.
Results: Upon univariate analysis, factors associated with 30-day complications were older age (P=0.03), non-white race, higher initial body-mass index, severe hypertension, diabetes mellitus, non-independent functional status and previous foregut/bariatric surgery (all P<0.001). These factors remained significant upon nominal logistic regression analysis (n=100,791, P<0.001, r2=0.008, AUROC=0.572). The ANN model is diagrammed in Figure 1A. Upon ANN analysis, the training set (80% of patients) was more accurate than logistic regression (n=80,633, r2=0.011, AUROC=0.581; Figure 1B), and confirmed by the validation set (n=20,158, r2=0.012, AUROC=0.585).
Conclusions: This study identifies a panel of simple and easily obtainable preoperative patient factors that may portend increased morbidity after LSG. Using an ANN model, prediction of these events can be optimized relative to standard logistic regression modeling.
Presented at the SAGES 2017 Annual Meeting in Houston, TX.
Abstract ID: 96044
Program Number: S053
Presentation Session: Bariatric I – Complications
Presentation Type: Podium