Fig. 4: Cut off selection to map 48 h mortality risk across the ICU stay and identification of features associated with change.

a Threshold plot showing Youden’s index, F-score, alarm rate, sensitivity, and specificity to differentiate three risk categories for predicted 48 h mortality. The lower threshold was selected based on the maximum Youden’s index, while the upper threshold was selected based on the maximum F-score. b Line plot illustrating ICU mortality based on fraction of ICU days in high-risk category per patient stay. Mortality was calculated from weighted density curves that were calculated by Kernel Density Estimation for patients with at least one day in the high-risk category. c Line plot depicting 48 h mortality risk calculated by the LGBM-48 algorithm for each day of the ICU stay of two patients (survivor, dashed line; non-survivor, dotted line). Indicators d and e refer to Figs. 4d and 4e that illustrate the change in SHAP values as 48 h mortality risk increases in the patient that succumbed. Horizontal solid lines identify lower and upper 48 h mortality risk thresholds that differentiate low, intermediate, and high risk. d,e calculated differences in SHAP values between the third-to-second (d) and second-to-last (e) 24 h interval before death for the non-surviving patient shown in Fig. 4c. The differences of SHAP value compared to SHAP values alone identifies shifts in feature importance affecting mortality risk predicted by the LGBM-48 h algorithm.