Bowels, meet brains: diagnosing IBS using deep learning approaches

Abstract

Irritable Bowel Syndrome (IBS) is a complex gastrointestinal disorder with a wide range of symptoms and thus no standardised diagnostic test, resulting in frequent misdiagnosis. This study explores the application of deep learning (DL) and machine learning (ML) techniques to classify IBS and its subtypes using multi-omics data, in particular, microbiome and metabolomics profiles.

Fecal metagenomic and metabolomic data were sourced from a publicly available dataset, preprocessed, and analysed to assess taxonomic and metabolic diversity between IBS patients and healthy controls. Alpha and beta diversity metrics showed moderate microbial variation, while metabolomic profiles revealed significant differences in key metabolites. Feature selection identified biomarkers such as xylose, salicylate, and thiamin (Vitamin B1) as significant indicators of IBS.

Several ML classifiers were evaluated for their ability to distinguish IBS from non-IBS and to subtype IBS cases. Among all models, an Artificial Neural Network (ANN) outperformed traditional approaches with a binary classification AUC of over 0.90 and a multiclass subtype classification accuracy of 75.54%. SHAP analysis provided interpretability of the ANN's predictions, linking biological relevance to computational outputs.

This research demonstrates the effectiveness of deep learning in leveraging metabolomic data for IBS diagnosis and subtype classification. It also highlights the potential of AI-driven clinical decision support systems for improving diagnostic accuracy and personalised treatment strategies in gastrointestinal health.