Development and validation of a machine learning scoring system based on multi-dimensional metabolic signatures for invasive examination triage in intestinal diseases.

BACKGROUND

Inflammatory bowel disease (IBD) and irritable bowel syndrome (IBS) often present with overlapping gastrointestinal symptoms despite distinct pathophysiological mechanisms. Colonoscopy remains the diagnostic gold standard but is invasive, costly and frequently overutilized. Single biomarkers have limited diagnostic value because they do not fully reflect the complex metabolic and inflammatory alterations underlying intestinal diseases. We aimed to develop and validate an interpretable machine learning-based scoring system integrating multidimensional metabolism-related biomarkers for invasive examination triage.

METHODS

This retrospective single-center study included 729 participants (313 healthy controls, 210 IBS, 100 ulcerative colitis and 106 Crohn's disease patients) enrolled between July 2021 and November 2025. Demographic, clinical, and metabolism-related laboratory biomarkers reflecting inflammatory metabolism, nutritional metabolism, hepatic metabolic function, and renal metabolic homeostasis were collected. Correlation network analysis, least absolute shrinkage and selection operator (LASSO) regression, extreme gradient boosting (XGBoost), and a simplified nomogram-based scoring system were applied to differentiate IBD from non-IBD conditions.

RESULTS

Significant differences were observed across all clinical and metabolism-related variables (P < 0.001). IBD patients exhibited elevated inflammatory-metabolic-related biomarkers and dense inflammation-driven metabolic correlation networks, whereas IBS patients showed metabolic profiles similar to healthy controls. XGBoost achieved the best diagnostic performance (AUC = 0.992), followed by LASSO (AUC = 0.978). A simplified scoring system incorporating sex, age, log-transformed fecal calprotectin (LogFC), log-transformed C-reactive protein (LogCRP), hemoglobin (HB), albumin (ALB), white blood cell count (WBC) and platelet count (PLT) achieved an AUC of 0.910, outperforming fecal calprotectin (FC) alone (AUC = 0.844) and the FC-CRP combination (AUC = 0.855). The scoring system demonstrated substantial clinical net benefit and correlated with colonoscopic inflammation severity (AUC = 0.761). Inflammatory and metabolism-related biomarkers were the strongest predictors of IBD.

CONCLUSION

IBD is characterized by distinct multidimensional metabolism-related biomarker signatures. The proposed machine learning-based scoring system integrates these metabolism-related biomarkers into a reliable, non-invasive tool for invasive examination triage, potentially reducing unnecessary colonoscopies and improving clinical resource utilization.