Machine learning approaches to the human metabolome in sepsis identify metabolic links with survival

Abstract

Background: Metabolic predictors and potential mediators of survival in sepsis have been incompletely characterized. We examined whether machine learning (ML) tools applied to the human plasma metabolome could consistently identify and prioritize metabolites implicated in sepsis survivorship, and whether these methods improved upon conventional statistical approaches.

Methods: Plasma gas chromatography-liquid chromatography mass spectrometry quantified 411 metabolites measured ≤ 72 h of ICU admission in 60 patients with sepsis at a single center (Brigham and Women's Hospital, Boston, USA). Seven ML approaches were trained to differentiate survivors from non-survivors. Model performance predicting 28 day mortality was assessed through internal cross-validation, and innate top-feature (metabolite) selection and rankings were compared across the 7 ML approaches and with conventional statistical methods (logistic regression). Metabolites were consensus ranked by a summary, ensemble ML ranking procedure weighing their contribution to mortality risk prediction across multiple ML models.

Results: Median (IQR) patient age was 58 (47, 62) years, 45% were women, and median (IQR) SOFA score was 9 (6, 12). Mortality at 28 days was 42%. The models' specificity ranged from 0.619 to 0.821. Partial least squares regression-discriminant analysis and nearest shrunken centroids prioritized the greatest number of metabolites identified by at least one other method. Penalized logistic regression demonstrated top-feature results that were consistent with many ML methods. Across the plasma metabolome, the 13 metabolites with the strongest linkage to mortality defined through an ensemble ML importance score included lactate, bilirubin, kynurenine, glycochenodeoxycholate, phenylalanine, and others. Four of these top 13 metabolites (3-hydroxyisobutyrate, indoleacetate, fucose, and glycolithocholate sulfate) have not been previously associated with sepsis survival. Many of the prioritized metabolites are constituents of the tryptophan, pyruvate, phenylalanine, pentose phosphate, and bile acid pathways.

Conclusions: We identified metabolites linked with sepsis survival, some confirming prior observations, and others representing new associations. The application of ensemble ML feature-ranking tools to metabolomic data may represent a promising statistical platform to support biologic target discovery.

Keywords: Artificial intelligence; Machine learning; Metabolism; Metabolomics; Sepsis.

Fig. 1

Heatmap showing normalized metabolite levels grouped by super-pathway among patients with sepsis. Heatmap showing normalized metabolite levels (rows), grouped by super-pathway, among patients with sepsis following hierarchical clustering (dendrogram, top); survival status is annotated (grey = yes, black = no). Differences in the metabolome among survivors and non-survivors were noted across multiple metabolic super-pathways

Fig. 2

Comparison of top metabolites selected by each analysis method’s innate feature selection algorithm. Comparison of top metabolites selected by each analysis method’s innate feature selection algorithm, identifying metabolites that more meaningfully contribute to successful sepsis mortality prediction models. Such approaches may identify measures of association and individual metabolic links with mortality. Agreement was noted between the lists of top metabolites identified by several machine learning methods, which also overlapped with those identified by conventional panelized logistic regression. FDA flexible discriminant analysis, GBM generalized boosted regression models, LR logistic regression, NSC nearest shrunken centroids, PLS-DA partial least squares-discriminant analysis, sparse LDA sparse discriminant analysis