Fig. 5.

Shapley analysis for variable importance and explanation of predictions. Panel A shows the mean absolute Shapley (SHAP) values averaged over all datasets (error bars indicate standard deviation over datasets). The top 20 variables are displayed. Large values indicate large contributions to the model’s prediction of sepsis. In the subpanel on the right, Shapley value distributions are exemplified for the eICU dataset. Positive Shapley values are indicative of positive predictions of the system and vice versa. In Panel B, Shapley values of individual heart rate and mean arterial pressure (MAP) measurements are shown. Panel C shows the Shapley values of the individual feature groups averaged across all datasets, whereas the available components of SOFA (labs and vitals), the count of lactate measurements, indication of oxygenation and ventilation as well as raw MAP values were most informative.