Machine Learning Framework for the Prediction of Severity and Mental Stress in IBD 2822

• Published in: The American journal of gastroenterology Vol. 113; no. Supplement; p. S1566
• Main Authors: Alvi, Azva, Friedman, Jacob, Won, Daehan, Bajwa, Maria, Bank, Leslie, Begdache, Lina
• Format: Journal Article
• Language: English
• Published: New York Wolters Kluwer Health Medical Research, Lippincott Williams & Wilkins 01.10.2018
• Subjects: Accuracy; Electronic health records; Gastroenterology; Machine learning; Mental disorders
• ISSN: 0002-9270; 1572-0241
• DOI: 10.14309/00000434-201810001-02821

Introduction: Irritable bowel disease (IBD) is a collection of inflammatory disorders of the digestive tract and colon. In terms of the severity of the disease, many investigations exist that attempt to explain dependency between attribute factors for patients but none include well established methods to determine a well-defined relationship while providing accurate predictive results based on the various factors. Furthermore, patients with IBD might be more sensitive to mental stress but there is still unclear about which factors are actually brining about the mental stress and how much they are affecting. To address those questions, we applied machine learning techniques to build predictive models for disease severity and existence of mental stress based on electronic health records (EHRs) and determined the relatively important factors to distinguish patients in terms of severity and the occurrence of mental disorder respectively. Methods: 697 patients (mean age: 49.33 years) EHRS were collected from UHS. We then selected 19 variables out of 42 unique variables including BMI, age, gender and so on, which were regarded as relatively significant to our target variables (i.e., responses; severity and mental stress). To build reliable models avoiding potential bias, the data imbalance (ex., the ratio of severity vs. non-severity "1:3.8) was resolved by a SMOTE (Synthetic Minority Over-Sampling Technique) algorithm. We considered four popular machine learning techniques: Support Vector Machine (SVM), Logistic Regression (LR), Decision Tree (DT), and Random Forest (RF). For the factors' importance measurement, stepwise and mean decrease methods were employed. Results: The RF model outperformed the other techniques while providing 93% accuracy for severity prediction and 81% accuracy for mental stress prediction respectively. In addition, the top five important variables to predict severity were the number of flare ups, BMI, age, family history, and presence of intestinal symptoms respectively. Conclusion: Severity level and the occurrence of mental disorder with IBD patients were distinguishable and predictable with a high degree of accuracy and reliability by use of only a few significant factors. The defined relationships are worth to investigate further for providing insights to understand underlying mechanism of the disease and will be beneficial for the clinical practice.