Figure 1.
Classification accuracies and the lung bacteria and untargeted metabolomics features (hydrophobic fraction) most strongly associated with each clinical outcome, as modeled in Framework 1. Results from DIABLO feature selection followed by elastic net models adjusted for age, sex, race, current smoking, inhaled corticosteroid use, and recent antibiotic use. (A) Mean out-of-sample classification accuracies. Red dashed line = 50% accuracy (random chance). Asterisks indicate mean model performance > random chance (one-sided t test). (B) Most predictive microbial and untargeted metabolomic features from adjusted elastic net models for outcomes whose classification accuracy exceeded random chance. Bacterial OTUs are displayed alphabetically, and metabolites are displayed by class membership with superclasses of interest indicated. See Table E5 for full IDs and class information. Metabolite names with >50 characters were relabeled as “Class-Name-#”. Superclasses are intended to highlight metabolite groups of interest; in particular, lipids. “Other” refers to compounds for which the superclass was unknown or the metabolite name, as displayed, provides indication of metabolite class. BDR = bronchodilator response; CAT = COPD Assessment Test; COPD = chronic obstructive pulmonary disease; FEF25–75 = maximum midexpiratory flow; HCU+AB/S = exacerbation requiring healthcare utilization and antibiotics/steroid treatment; ns = not significant; OTU = operational taxonomic unit; SGRQ = St. George’s Respiratory Questionnaire.