(393) Machine Learning Classifier Models Predict Metabolic Phenotypes: post-hoc study of VITdAL-ICU trial

Introduction: Vitamin D deficiency is common in critically ill patients and associated with worse outcomes. High-dose vitamin D3 supplementation trials have failed to show benefits, but some potential beneficial subgroups have been identified in prior studies. This study aims to predict metabolomic sub-phenotypes in critically ill patients with vitamin D deficiency by machine learning classifier models with reduced metabolite data.

Methods: As a post-hoc analysis of the VITdAL-ICU trial [the Correction of Vitamin D Deficiency in Critically Ill Patients], we analyzed 659 metabolites from 453 patients at randomization. We applied the random forest classifier model and gradient boost classifier model to predict prior-identified four metabolomic sub-phenotypes in the divided training data. Secondly, we performed additional models with decreased metabolites as predictors based on their importance. Finally, we evaluated the predictivity of the models with fewer metabolites in the separate test data.

Results: In the primary analysis, classifier models demonstrated high accuracy in the train set [the random forest model: 81% (95% CI, 0.73–0.87), the gradient boost model: 80% (95% CI, 0.73–0.86)]. We identified seven high-yield metabolites; 2-methylbutyroylcarnitine (C5), S-adenosylhomocysteine, hexadecanedioate (C16), 1-lignoceroyl-GPC (24:0), biliverdin, trigonelline (N’-methylnicotinate), and methyl indole-3-acetate. As evaluation with the test dataset, classifier models with 7 metabolites held high predictivity [the random forest model: AUC of 0.91 (95% CI, 0.85–0.95); the gradient boost model: AUC of 0.88 (95% CI, 0.80–0.95)]. Finally, the treatment interaction was significant with assigned phenotype D with vitamin D supplementation regarding 180-day mortality in the random forest model(p = 0.02).

Conclusions: Metabolic sub-phenotypes in critical illness with vitamin D deficiency can be accurately classified by machine learning models based on high-yield metabolites. Further validation studies will be required to generalize our findings to critically ill populations.