Fig. 3: Performance of the selected LGBM-48h model and feature importance. | Communications Medicine

Fig. 3: Performance of the selected LGBM-48h model and feature importance.

From: An interpretable machine learning algorithm enables dynamic 48-hour mortality prediction during an ICU stay

Fig. 3

a Receiver operating characteristic curves illustrating the predictive performance for 48 h ICU mortality of the LGBM-48 h algorithm, the SAPS II/TISS-10 and the SOFA score. b Forest plot illustrating the predictive performance for 48 h ICU mortality of the LGBM-48 h algorithm, stratified by stay day after ICU admission. Black filled circles represent individual AUROC values for each ICU stay day after admission. Error bars indicate 95% confidence intervals derived using the DeLong method. Dotted line indicates area under the receiver operating characteristic curve (AUROC) for the LGBM-48 h algorithm, shaded blue area illustrate 95% confidence intervals. Gray bars on the right indicate number of stay days that were included in the analysis (n = 30,171 stay days in the test set). c Violin and dot plot illustrating the 15 features with the highest mean absolute SHAP values. The bar chart on the right summarizes the mean absolute SHAP value for each feature. d Line graph illustrating the change of mean absolute SHAP values for seven features with the highest absolute SHAP values at time of death. In patients that succumb, mean absolute SHAP values change in proximity to death.

Back to article page