Machine Learning Techniques to Predict SIBO Breath-Test Positivity Based on Presenting Symptoms

Introduction: Small intestinal bacterial overgrowth (SIBO), defined as the presence of excessive bacteria in the small bowel, has been associated with a wide range of complaints including abdominal bloating, flatulence, weight loss and diarrhea and is usually diagnosed on the basis of a hydrogen and methane breath test. The purpose of our study was to determine which among these symptoms were predictive of breath-test positivity. Methods: Data was retrospectively collected on patients receiving a SIBO breath test over a 3-year period from 2018 to 2021. Information was gathered on patient demographics as well as on the nature of the symptoms that prompted performance of the breath test. A chi-squared analysis was conducted on each presenting symptom in relation to SIBO positivity. A multivariate analysis was then conducted to determine if complex relationships between presenting symptoms exist and could predict SIBO positivity. A logistic regression (LR) was first used for evaluation followed by a gradient boosted decision tree (GBDT) and then a logistic regression boosted decision tree (LR+GBDT). The models were hyptertuned, then evaluated with k-fold cross validation. Resulting AUCs are reported along with Gini importance when relevant. Results: A total of 390 patients with SIBO tests were collected. One hundred seventy-four were positive and 216 negative. On univariate analysis neither the presence of bloating, distension, nausea, vomiting, flatulence, pain, discomfort, diarrhea, steatorrhea, weight loss or constipation significantly correlated with SIBO positivity. Logistic regression after 10-fold cross validation yielded an AUC of 0.44 ± 0.09. GBDT yielded an AUC of 0.46 ± 0.07 and LR+GBDT 0.46 ± 0.09. Gini importance of the LR+GBDT had age as the largest contributor at 0.26 and female sex next at 0.06 (Figure 1). Conclusion: In general, it appears that patient reported symptoms are poor predicters of SIBO breath test positivity. This remained true even with the implementation of various machine learning techniques. Age was the best predictor of SIBO-breath test positivity. This is likely secondary to age-related decline in gastric acid secretion and GI motility. Elderly patients are also more likely to be prescribed GI motility altering medications and have comorbid conditions which impact gut motor function.Figure 1.: Gini Importances from Logistic Regression and Gradient-Boosted Tree Classifier.