Fig. 4

GLM with feature selection provides improved DAS28-CRP prediction accuracy in an independent validation group (12 samples). A Performance of GLMs in predicting quantitative disease activity was evaluated on samples of an independent validation group. Distributions of absolute errors from models with and without a feature selection scheme were compared to identify the more robust model. B Selection of metabolic features prior to model training resulted in higher predictive performance, as evidenced by the stronger correlation between observed and predicted DAS28-CRPs. Three samples predicted to have negative DAS28-CRP values are omitted from the scatter plot. The dashed violet line indicates “y = x,” i.e., an exact match between the observed and predicted values. 95% confidence interval for ρ with feature selection [0.18, 0.90]; without feature selection [−0.44, 0.68]