Classifying irritable bowel syndrome using spatio-temporal graph convolution networks on brain functional MRI data

Irritable bowel syndrome (IBS) is a functional gastrointestinal disorder marked by abdominal pain and changes in stool consistency or frequency. Recent studies have explored the link between IBS and alterations in brain networks using functional MRI. Despite these efforts, an effective diagnostic or predictive model for IBS remains elusive. This shortfall is twofold: firstly, the sample sizes in these studies are typically small, and secondly, the machine learning or deep learning models currently in use fail to adequately detect the subtle and dynamic pathological changes present in functional MRI data for IBS. In this study, we extracted rs-functional MRI of 79 subjects with IBS and 79 healthy controls, then put them into spatio-temporal graph convolution network (ST-GCN) for classification. We also incorporated a novel interpretability module into this model to identify potential regions of interest (ROI) associated with IBS. Our model outperformed other state-of-the-art machine learning and deep learning methods with the highest average accuracy of 83.51% on our dataset. Furthermore, based on the results of our interpretability module, the Inferior Parietal Lobule (IPL.R), Inferior Frontal Orbital part (ORBinf.R), Postcentral Gyrus (PCG.R), Middle Frontal Orbital part (ORBmid.R), and Superior Medial Frontal Orbital part (ORBsupmed.L) were identified as top 5 important brain regions for distinguishing IBS patients from the control group, which are consistent with the brain regions identified in previous literature reviews. We also conducted three external data-driven experiments to further validate the effectiveness of the interpretability module: (1) Experiment only on important brain regions; (2) Comparison with the Perturbation-Based Methods; (3) Correlation analysis. The results indicate that the selected regions significantly impact IBS.