Fig. 4. Prediction of TBI patient outcomes.

a The ROC curves and AUC values of four penalized logistic regression models. Lasso logistic regression and ridge logistic regression were evaluated with two sets of features each. The first set of features was the full metabolomics dataset (459 features). The second set of predictors was the top features as selected by random forest feature selection and Welch t testing (19 features). The curves and AUC values are the average of 100 training/testing folds. b Individual discriminatory performance for the top 19 metabolites. Each metabolite was used in a logistic regression model as predictor, with outcome as response. The performance was averaged on 100 model runs of 70–30% data splits. Data are presented as mean values with the individual run performances as points (n = 100) and aggregated 95% CI. c Heatmap of the top 19 features (also used in the reduced models in panel a), as selected by the random forest and the Welch t test feature selection. GOSe of 1–4 is considered as unfavorable outcome and GOSe of 5–8 as favorable. Overall, patients with favorable outcomes have lower concentration of metabolite/lipid levels, with a notable exception of Glycerol, which is in higher levels in patients with favorable outcomes. d Evaluation of the discriminatory performance of logistic regression models for different cut-offs of GOSe values (1 vs. 2–8, 1–2 vs. 3–8, …. 1–7 vs. 8). The AUC (red points) and CI values are the average of 100 training/testing folds for each cut-off and each severity group. It appears that the accurate discrimination of full recovery (GOSe of 8) is not possible with the metabolomic/lipid dataset. e Pathway analysis using MetaboAnalyst³¹ tool. The enriched metabolic pathways are based on differences of serum metabolites between the favorable and unfavorable outcome groups. Only significantly different pathways (FDR corrected p-values from t test) with 2 or more hits are included.