Assessment of Self-Reported Executive Function in Patients with Irritable Bowel Syndrome Using a Machine-Learning Framework

Abstract

Introduction: Irritable bowel syndrome (IBS) is characterized as a disorder of the gut-brain interaction (DGBI). Here, we explored the presence of problems related to executive function (EF) in patients with IBS and tested the relative importance of cognitive features involved in EF. Methods: A total of 44 patients with IBS and 22 healthy controls (HCs) completed the Behavior Rating Inventory of Executive Function (BRIEF-A), used to identify nine EF features. The PyCaret 3.0 machine-learning library in Python was used to explore the data, generate a robust model to classify patients with IBS versus HCs and identify the relative importance of the EF features in this model. The robustness of the model was evaluated by training the model on a subset of data and testing it on the unseen, hold-out dataset. Results: The explorative analysis showed that patients with IBS reported significantly more severe EF problems than the HC group on measures of working memory function, initiation, cognitive flexibility and emotional control. Impairment at a level in need of clinical attention was found in up to 40% on some of these scales. When the nine EF features were used as input to a collection of different binary classifiers, the Extreme Gradient Boosting algorithm (XGBoost) showed superior performance. The working memory subscale was consistently selected with the strongest importance in this model, followed by planning and emotional control. The goodness of the machine-learning model was confirmed in an unseen dataset by correctly classifying 85% of the IBS patients. Conclusions: The results showed the presence of EF-related problems in patients with IBS, with a substantial impact of problems related to working memory function. These results suggest that EF should be part of an assessment procedure when a patient presents other symptoms of IBS and that working memory function should be considered a target when treating patients with the disorder. Further studies should include measures of EF as part of the symptom cluster characterizing patients with IBS and other DGBIs.

Keywords: BRIEF-A; IBS; executive function; feature importance; gut–brain axis; machine learning.

Figure 1

(a) The overlapping age distributions of males and females in the IBS (n = 44) and the HC (n = 22) group. (b) Distribution of IBS-SSS scores in the IBS (n = 40) and the HC (n = 17) group. The vertical dashed line denotes the mean IBS-SSS score in each group.

Figure 2

Pairwise scatterplots between all BRIEF-A subscales color-coded separately for the IBS and the HC group. The distributions are fitted with a least squares regression line with a shaded confidence interval. The diagonal entries show the group-specific kernel density estimated distributions for each BRIEF-A subscore.

Figure 3

The mean metrics across each of the 10 folds in the training set. The ranking is defined with respect to mean Accuracy across the folds where the mean performance measures of the other metrics (AUC, Recall, Precision, F1, Kappa and MCC) for the different classifiers provided by PyCaret (left column) are given. Each classifier is applied on the same sets of folds, generated randomly from the training set, in our 10-fold cross-validation scheme. The most prominent mean values among the performance metrics are highlighted in yellow. Due to its overall best performance, we selected the XGboost classifier in the following analysis. Note, the dummy classifier makes predictions that ignore the input features, e.g., it returns the most frequent class label, serving as a baseline to compare against more complex classifiers.

Figure 4

(a) Ranking of the permutation feature importance computed from the training set. (b) Ranking of mean absolute SHAP values computed in the training set. Note that the top three rankings are the same regarding order and qualitative differences for both the permutation importance method and in the SHAP values from cooperative game theory.

Figure 5

Partial dependence plots (PDP) and individual conditional expectation (ICE) plots for each of the nine BRIEF-A subscales. The collection ($n=46$) of thin traces across the range of BRIEF-A values, one line per subject in the training set, shows how the subject’s prediction changes when a feature (BRIEF-A subscale) changes (typically within the interval 30 to 80). Note that some of the traces are visually inseparable. The predicted outcome on the vertical axes denotes a continuous scale between 0 (=HC) and 1 (=IBS).

Figure 6

(a) Confusion matrix for the binary XGBoost classification of IBS versus HC based on the 9-dimensional feature vectors from the BRIEF domains. (b) BRIEF item-wise percentage of clinically impaired participants in the IBS and HC group, i.e., percentage of BRIEF variables $\ge \mathrm{cutoff}$ (=65) in the IBS (n = 44) and HC (n = 22) groups.