Fig. 4: Feature inspection.
From: Early prediction of circulatory failure in the intensive care unit using machine learning
a, The 15 features with highest mean absolute SHAP values. On the y axis, the black violin plot shows the full distribution of the SHAP values for each feature. The dot plot in the foreground shows a color coding of the actual value of the feature, resulting in the SHAP value as indicated on the x axis. The color coding is based on the percentile of the feature value with respect to the whole distribution. b,c, Scatter plots showing the relationship between feature value and SHAP value for the minimal MAP in the last hour (b) and for patient age (c). The orange line and shade represents the mean and s.d. of the regression line. The distributions of the SHAP and feature values are shown as histograms on the right and top of the scatter graph. The high variance in the SHAP value for a given feature value indicates a strong influence of other features. d, Precision-recall curves illustrating the performance of the circEWS-lite model on the original test set (Original) and on the resampled test set with regular sampling intervals (Resampled), as well as the performance of the circEWS-lite model trained on binarized data in the binarized test set (Binarized). The variation estimate was derived from n = 5 independent experiments in the temporal splits. e, Shapelet feature illustration. The lactate values are shown over the indicated time since admission. The lactate shapelet shows an increase in its SHAP value (gray line) 5.5 h before the patient suffers from a circulatory deterioration (blue arrow). The light blue line indicates the feature value that represents the L2 distance between the time series and the shapelet at a 4-h delay. The feature value drops right before the event, increasing the prediction score (red arrow).
