Machine learning-based phenogroups and prediction model in patients with functional gastrointestinal disorders to reveal distinct disease subsets associated with gas production.

Abstract

Symptom-based subtyping for functional gastrointestinal disorders (FGIDs) has limited value in identifying underlying mechanisms and guiding therapeutic strategies. Small intestinal dysbiosis is implicated in the development of FGIDs. We tested if machine learning (ML) algorithms utilizing both gastrointestinal (GI) symptom characteristics and lactulose breath tests could provide distinct clusters. This was a prospective cohort study. We performed lactulose hydrogen methane breath tests and hydrogen sulfide breath tests in 508 patients with GI symptoms. An unsupervised ML algorithm was used to categorize subjects by integrating GI symptoms and breath gas characteristics. Generalized Estimating Equation (GEE) models were used to examine the longitudinal associations between cluster patterns and breath gas time profiles. An ML-based prediction model for identifying excessive gas production in FGIDs patients was developed and internal validation was performed. FGIDs were confirmed in 300 patients. K-means clustering identified 4 distinct clusters. Cluster 2, 3, and 4 showed enrichments for abdominal distention and diarrhea with a high proportion of excessive gas production, whereas Cluster 1 was characterized by moderate lower abdominal discomforts with the most psychological complaints and the lowest proportion of excessive gas production. GEE models showed that breath gas concentrations varied among different clusters over time. We further sought to develop an ML-based prediction model to determine excessive gas production. The model exhibited good predictive capabilities. ML-based phenogroups and prediction model approaches could provide distinct FGIDs subsets and efficiently determine FGIDs subsets with greater gas production, thereby facilitating clinical decision-making and guiding treatment.