A psychological symptom based machine learning model for clinical evaluation of irritable bowel syndrome

Abstract

Background: Irritable bowel syndrome (IBS) is a chronic functional gastrointestinal disorder characterized by recurrent abdominal pain associated with alterations in stool form and/or stool frequency. Co-morbidities such as anxiety, depression, fatigue, and insomnia are frequently reported by patients suffering from IBS. Identification of these symptoms should thus be an integral part of an IBS assessment. However, an optimal tool to screen for core psychological symptoms in IBS is still missing. Here, we aim to develop a psychological symptom based machine learning model to efficiently help clinicians to identify patients suffering from IBS. Methods: We developed a machine learning workflow to select the most significant psychological features associated with IBS in a dataset including 49 patients with IBS and 35 healthy controls. These features were used to train three different types of machine learning models: logistic regression, decision trees and support vector machine classifiers; which were validated on a holdout validation dataset and an unseen test set. The performance of these models was compared in terms of balanced accuracy scores. Results: A logistic regression model including a combination of symptom features associated with anxiety and fatigue resulted in a balanced accuracy score of 0.93 (0.81-1.0) on unseen test data and outperformed the other comparable models. The same model correctly identified all patients with IBS in a test set (recall score 1) and misclassified one non-IBS subject (precision score 0.91). A complementary post-hoc leave-one-out cross validation analysis including the same symptom features showed similar, but slightly inferior results (balanced accuracy 0.84, recall 0.88, precision 0.86). Conclusions: Inclusion of machine learning based psychological evaluation can complement and improve existing clinical procedure for diagnosis of IBS.

Keywords: Irritable bowel syndrome; anxiety; classification; fatigue; machine learning.

Figure 1.. Various pools of features generated by arranging four questionnaires into all possible combinations.

ANX = anxiety questionnaire; DEP = depression questionnaire; BIS = Bergen Insomnia Scale; CFS = Chalder Fatigue Scale.

Figure 2.. Sequential backward feature selection from input feature set.

Figure 3.. Classification of IBS and HC in the unseen test dataset (n = 17) using three different machine learning models.

IBS = Participants with irritable bowel syndrome; HC = Healthy control subjects.

Figure 4.. Feature contribution for the logistic regression model using permutation feature importance.

CFS_Q1. = First question in Chalder Fatigue Scale; CFS_Q6 = Sixth question in Chalder Fatigue Scale; ANX_Q1 = First question in anxiety questionnaire; Gender = Participant gender; CFS_Q12 = Twelfth question in Chalder Fatigue Scale; CFS_Q8 = Eighth question in Chalder Fatigue Scale; ANX_Q3 = Third question in anxiety questionnaire; ANX_Q2 = Second question in anxiety questionnaire; CFS_Q5 = Fifth question in Chalder Fatigue Scale.