Fig. 3: Risk profiles for selected top features. | Communications Medicine

Fig. 3: Risk profiles for selected top features.

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

Fig. 3

a FASLG and (b) GDF15 have similar feature importance in the EBM Proteomics & Clinical model, but the risk profiles differ drastically. The two clinical features (c) Cystatin C (field 30720) and d None of the medications* (field 6153, value −7) in the EBM Clinical model with similar risk profiles show drastically different risk profiles and feature distributions. c A small number of participants with high Cystatin C levels show greatly increased CVD risk. d Participants who selected “none of the above” when reporting medication use* showed reduced risk of CVD. This is a weak effect affecting a large number of individuals. Risk profiles for (e) HDL (field 30780) and f MSR1 show high consistency. The top panels of each subfigure display the risk score on the logarithmic scale (y-axis) for different feature values (x-axis), with gray-shaded regions representing the standard deviation for estimations. The dotted line corresponds to the population mean risk for the given feature. Positive values indicate increased risk; negative values indicate decreased risk. The bottom panels show histograms of the feature values across all participants. *Medications for cholesterol, blood pressure, diabetes, or exogenous hormones.

Back to article page