Temporal interpretation of physiologic group contributions for ICU mortality prediction
- 1Department of Industrial Data Engineering, Hanyang University
- 2Department of Medical and Digital Engineering, Hanyang University
Abstract
Reliable ICU mortality prediction requires capturing physiologic evolution over time and interpreting these patterns in clinically meaningful ways. However, existing methods often focus on individual variables or isolated time points, limiting their ability to represent related physiologic states and temporal progression. We developed a group-based temporal interpretation framework by applying GroupSegment-SHAP to characterize the evolving contributions of physiologic groups before ICU mortality. Using MIMIC-IV, we evaluated prediction models across 24-, 48-, and 72-hour death-aligned lookback windows and selected Mamba, a selective state-space model, as the interpretation backbone because it showed comparatively strong and stable performance. Clinical variables were reorganized into physiologic groups, and Shapley values from the trained Mamba model were used to quantify group-wise temporal importance. Group-level interpretation identified hemodynamic, neurologic, cardiopulmonary, and temperature-related groups as major mortality predictors. Temporal analysis showed that 24-hour trajectories emphasized acute cardiopulmonary, metabolic, hematologic, and acid-base/respiratory abnormalities, whereas 48- and 72-hour trajectories additionally captured hemodynamic, renal-output, neurologic, and hepatic differences. Cardiopulmonary importance increased 8 to 30 h before peak group-level importance, when non-survivors showed lower normal-range proportions than survivors by up to 17.5% points. This analysis shifts interpretation from static variable-level importance to temporal trajectories of physiologic groups.
Related
- Machine learning prediction model for distant metastasis at diagnosis in pediatric neuroblastoma: development, interpretation, and temporal validation
- Temporal Validation and Simplification of Mortality Prediction for Perioperative Cardiopulmonary Resuscitation
- Incremental value of bilateral ankle-brachial index for ten-year heart and cerebrovascular mortality prediction: an exploratory temporal benchmark in NHANES
- Temporal Trends and Demographic Contributions to Stroke Burden in India, 1990-2023: An Analysis of Global Burden of Disease 2023 Estimates Using Shapley Decomposition and Joinpoint Regression.
- Synergistic contributions of periodic and aperiodic cortical activity to motor recovery prediction after stroke
BibTeX
@article{kim2026temporal,
title = {{Temporal interpretation of physiologic group contributions for ICU mortality prediction}},
author = {Jinwoong Kim and Yeeun Kim and Sangjin Park},
journal = {Scientific Reports},
year = {2026},
month = jul,
doi = {10.1038/s41598-026-62826-3},
url = {https://doi.org/10.1038/s41598-026-62826-3},
note = {CC BY},
}
RIS
TY - JOUR
TI - Temporal interpretation of physiologic group contributions for ICU mortality prediction
AU - Jinwoong Kim
AU - Yeeun Kim
AU - Sangjin Park
PY - 2026
DA - 2026/07/01
JO - Scientific Reports
DO - 10.1038/s41598-026-62826-3
UR - https://doi.org/10.1038/s41598-026-62826-3
AB - Reliable ICU mortality prediction requires capturing physiologic evolution over time and interpreting these patterns in clinically meaningful ways. However, existing methods often focus on individual variables or isolated time points, limiting their ability to represent related physiologic states and temporal progression. We developed a group-based temporal interpretation framework by applying GroupSegment-SHAP to characterize the evolving contributions of physiologic groups before ICU mortality. Using MIMIC-IV, we evaluated prediction models across 24-, 48-, and 72-hour death-aligned lookback windows and selected Mamba, a selective state-space model, as the interpretation backbone because it showed comparatively strong and stable performance. Clinical variables were reorganized into physiologic groups, and Shapley values from the trained Mamba model were used to quantify group-wise temporal importance. Group-level interpretation identified hemodynamic, neurologic, cardiopulmonary, and temperature-related groups as major mortality predictors. Temporal analysis showed that 24-hour trajectories emphasized acute cardiopulmonary, metabolic, hematologic, and acid-base/respiratory abnormalities, whereas 48- and 72-hour trajectories additionally captured hemodynamic, renal-output, neurologic, and hepatic differences. Cardiopulmonary importance increased 8 to 30 h before peak group-level importance, when non-survivors showed lower normal-range proportions than survivors by up to 17.5% points. This analysis shifts interpretation from static variable-level importance to temporal trajectories of physiologic groups.
SN - 2045-2322
ER -