S.C. Perez1, L. Coghill2, A.A. Wheeler1 1University Of Missouri, Surgery, Columbia, MO, USA 2University Of Missouri, Bioinformatics, Columbia, MO, USA
Introduction: Postoperative small bowel obstruction (SBO) is a rare postoperative complication that occurs after primary bariatric surgery. Currently, there are no adequate predictive models to determine the risk of SBO, however, machine learning techniques such as deep neural networks (a primary component of artificial intelligence) and extreme gradient boosting (XGB) potentially offer a substantial advantage over conventional prediction methods. Therefore, the primary aim of this study is to determine the utility of more advanced risk prediction techniques in the field of surgery.
Methods: 729,050 patients from the MBSAQIP database were analyzed, and patients undergoing primary laparoscopic gastric bypass (GB) or sleeve gastrectomy (SG) were included. The machine learning models used were logistic regression (LR), DNN and XGB to predict postoperative SBO based on available preoperative and intraoperative variables captured by the MBSAQIP database.
Results: The DNN and XGB models outperformed the LR model with AUC values of 0.832 95%CI (0.814-0.851), 0.829 95%CI (0.810-0.848), and 0.768 95%CI (0.747-0.788) respectively. After threshold optimization for each prediction method, DNN and XGB methods outperformed conventional LR with sensitivity values of 0.800 and 0.790 versus 0.687 respectively. Specificity values for the DNN, XGB, and LR models were 0.756, 0.777, and 0.707. The top five variables with the most important influence on the predictive outcome, according to the XGB and LG outputs were procedure type, operation length, creatinine level, age, and weight. Additional variables with mild importance regarding model strength are operation year, race, preoperative hematocrit, and the number of hypertensive medications.
Conclusion: Machine learning methods are currently outperforming traditional predictive models in surgery and are underutilized in specialties with large databases available. Increased use and awareness of these techniques will help bring surgical risk prediction into the modern era while also providing patients and surgeons with more detailed risk/benefit analysis.