Fig. 2: EBM models provide local and global explanations for feature importance. | Communications Medicine

Fig. 2: EBM models provide local and global explanations for feature importance.

From: Interpretable machine learning leverages proteomics to improve cardiovascular disease risk prediction and biomarker identification

Fig. 2

a–c Local explanations quantify the contribution of different features to an individual participant’s predicted CVD risk. The ribbon of each panel represents the label and the probability of the label, as provided by the EBM Proteomics model. For instance, “Case, P(Case) = 0.564” means that the participant was indeed a case, and the model assessed (without knowledge of the outcome) that they had a 56.4% probability of being a case. The model quantifies the contribution of each protein to CVD risk for every participant. Red bars represent increased risks based on the plasma protein expression level, while blue bars denote reduced disease risks according to the expression level. Note participants in (a, b) show similar risk but different contributing proteins, while the participant in (c) shows low disease risk. The intercept, with a value of −3.04, is not shown. d Global explanations aggregate the contribution of features across the cohort. The graph displays the top 10 contributing proteins in the EBM model. Known CVD markers were highlighted in red.

Back to article page