J. D. Bozzay1,19, P. F. Walker1,19, V. Khatri1,17,19, M. Zielinski2, S. Wydo3, D. Cullinane4, J. Dunn5, T. Duane6, D. Turay7, K. Inaba8, R. Lesperance9, M. Rosenthal10, J. Watras11, A. Pakula12, K. A. Widom13, J. Cull14, E. Toschlog15, T. Z. Hayward16, S. Schobel-Mchugh1,17,19, E. A. Elster1,17,19, C. J. Rodriguez1,19, M. J. Bradley1,17,18,19 1Walter Reed National Military Medical Center,Department Of Surgery,Bethesda, MD, USA 2Mayo Clinic,Department Of Surgery,Rochester, MN, USA 3Cooper University Hospital,Department Of Surgery,Camden, NJ, USA 4Marshfield Clinic,Department Of Surgery,Marshfield, WI, USA 5UC Health Northern Colorado,Department Of Surgery,Loveland, CO, USA 6John Peter Smith,Department Of Surgery,Forth Worth, TX, USA 7Loma Linda University Health,Department Of Surgery,Loma LInda, CA, USA 8Keck School of Medicine of USC,Department Of Surgery,Los Angeles, CA, USA 9San Antonio Military Medical Center,Department Of Surgery,Fort Sam Houston, TX, USA 10Massachusetts General Hospital,Department Of Surgery,Boston, MA, USA 11Inova Fairfax Hospital,Department Of Surgery,Falls Church, VA, USA 12Kern Medical Center,Department Of Surgery,Bakersfield, CA, USA 13Geisinger Medical Center,Department Of Surgery,Danville, PA, USA 14Greenville Memorial Hospital,Department Of Surgery,Greenville, SC, USA 15East Carolina University,Department Of Surgery,Greenville, NC, USA 16Indiana University School Of Medicine,Department Of Surgery,Indianapolis, IN, USA 17Surgical Critical Care Initiative,Bethesda, MD, USA 18Naval Medical Research Center,Department Of Regenerative Medicine,Bethesda, MD, USA 19Uniformed Services University Of The Health Sciences,Bethesda, MD, USA
Introduction: Identifying candidates who will require therapeutic surgery (TS) for non-emergent small bowel obstruction (SBO) remains challenging. Machine learning models can elicit complex dependencies and may perform better than traditional regression models. The objective of this study was to compare both strategies to best identify patients who would require TS for the management of SBO.
Methods: A prospectively maintained multi-institutional database from the Eastern Association for the Surgery of Trauma was reviewed. Presence of peritonitis, closed loop obstruction on imaging, virgin abdomen, or patients with data paucity were excluded, leaving 566 patients for analysis. Random Forest (RF) and logistic regression (LR) models were generated separately for both gastrografin challenge (GC) and non-GC patients.
Results: 156 (27.6%) patients underwent TS. The non-GC RF model produced an area under the curve (AUC) of 0.68, sensitivity of 0.64, and specificity of 0.70. The non-GC LR model produced an AUC of 0.62, sensitivity of 0.59, and specificity of 0.65. The GC RF model produced an AUC of 0.89, sensitivity of 0.86, and specificity of 0.89. The GC LR model produced an AUC of 0.89, sensitivity of 0.87, and specificity of 0.87. Predictive variables for therapeutic surgical intervention for the GC RF and LR models included GC test result, systolic blood pressure, presence of intraperitoneal fluid, presence of CT transition point, and previous occurrence of at least of 1 of the following: Crohn’s disease, enterocutaneous fistula, gastric bypass, metastatic cancer, small bowel obstruction, or ventral hernia. In the GC RF and LR models, removal of the GC test result as a predictor, substantially lessened performance metrics for both the RF (AUC of 0.59, sensitivity of 0.57, specificity of 0.64) and LR models (AUC of 0.61, sensitivity of 0.62, specificity of 0.65). The GC test result alone had a sensitivity of 0.7 and specificity of 0.93.
Conclusion: An accurate model for predicting the need for SBO TS was developed using a combination of clinical and radiographic data. Furthermore, incorporation of the GC significantly improves model performance and is an important clinical test during the workup of non-emergent SBO. The improved performance for GC patients is critically dependent on the inclusion of GC result as a predictor. This type of predictive modeling may be a useful adjunct to support future clinical decision-making. Evaluation with an external validation dataset is required to assess the generalizability of model performance.
