Accurate assessment of enteral feeding intolerance (ENFI) in septic patients remains challenging, as existing clinical tools show limited predictive value.

A prospectively cohort of 60 patients with sepsis (30 ENFI, 30 with feeding tolerance) and 20 healthy controls was enrolled. Serum samples were drawn for metabolomic profiling on the 1st day following sepsis. LC/MS was used to profile serum metabolites. The feeding intolerance outcome was collected at 14 days after ICU admission. Using a deep learning algorithm, we developed three ENFI prediction models: a metabolite-based model, a clinical risk model, and a new combined model integrating both feature types.

The metabolite-based model included four key biomarkers—palmitic acid, histidine-threonine, glutamate-histidine, and dehydrobilirubin—and achieved an area under the receiver operating characteristic curve (AUC) of 0.85. The clinical model, based on variables such as APACHE II score, intra-abdominal pressure, and albumin, achieved an AUC of 0.88. The combined model demonstrated the best performance, with an AUC of 0.94. It also showed higher accuracy, precision, F1-score, and net benefit in decision curve analysis, particularly when the risk threshold exceeded 4%. Statistical comparisons confirmed its superiority (Net Reclassification Index = 0.33, Integrated Discrimination Improvement = 0.31, P < 0.05).

Integrating metabolomics with clinical data significantly improves ENFI risk prediction in septic patients. External validation is warranted before clinical application.