Towards Interpretable Risk: Multidimensional Context for ICU Mortality Predictions
- 1Institute for Population and Precision Health, University of Chicago, Chicago, IL, USA
Abstract
ICU mortality models can achieve strong discrimination, yet a risk score alone provides limited context for patient-level interpretation. We developed a multidimensional prediction-context framework that complements a calibrated mortality estimate with model behavior, data availability, recent physiology, and model attribution. Using 3,236 held-out ICU episodes from the MIMIC-III in-hospital mortality benchmark, we compared LSTM, GRU-D, XGBoost, and a weighted ensemble. The ensemble achieved an AUROC of 0.871 and AUPRC of 0.536; validation-based logistic recalibration improved the Brier score from 0.134 to 0.075. Incorrect predictions showed greater component-model disagreement and smaller decision-boundary margins, although model agreement and large margins did not guarantee correctness. Observation coverage and recent physiological trends also varied substantially across patients, highlighting differences in the information surrounding otherwise similar risk estimates. SHAP analysis attributed 89.9% of total absolute XGBoost attribution to physiological-value features and 10.1% to observation-process features. These dimensions were integrated into patient-level profiles to provide a more complete view of how predictions were formed and the clinical and data context surrounding them, extending interpretation beyond risk scores and feature rankings alone. Abbreviations: ICU (Intensive care unit)
Related
- Chronic kidney disease–mineral and bone disorder as multidimensional bone–vascular risk framework: from biochemical control to risk-and context-guided care
- An interpretable machine learning model for predicting intensive care unit mortality risk in elderly patients with concomitant cardiogenic shock and sepsis: Development and external validation using two large critical care databases
- Machine learning-based prediction of lymph node metastasis in papillary thyroid carcinoma: an interpretable multidimensional model
- XGBoost-Cox model effectively predicts mortality risk in elderly male patients with acute kidney injury: a retrospective cohort study integrating nonlinear feature engineering and interpretable machine learning
- Public health determinants of life expectancy in Saudi Arabia: a multidimensional analysis of infant mortality, anemia, immunization, fertility, and health expenditure
BibTeX
@misc{gupta2026towards,
title = {{Towards Interpretable Risk: Multidimensional Context for ICU Mortality Predictions}},
author = {S. Gupta and A. Das and M. S. Anto and Z. Alam and A. Datta and T. Gupta and T. S. Pias and H. Islam},
howpublished = {medRxiv},
year = {2026},
month = sep,
doi = {10.64898/2026.09.03.26362203},
url = {https://www.medrxiv.org/content/10.64898/2026.09.03.26362203v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Towards Interpretable Risk: Multidimensional Context for ICU Mortality Predictions
AU - S. Gupta
AU - A. Das
AU - M. S. Anto
AU - Z. Alam
AU - A. Datta
AU - T. Gupta
AU - T. S. Pias
AU - H. Islam
PY - 2026
DA - 2026/09/03
PB - medRxiv
DO - 10.64898/2026.09.03.26362203
UR - https://www.medrxiv.org/content/10.64898/2026.09.03.26362203v1
AB - ICU mortality models can achieve strong discrimination, yet a risk score alone provides limited context for patient-level interpretation. We developed a multidimensional prediction-context framework that complements a calibrated mortality estimate with model behavior, data availability, recent physiology, and model attribution. Using 3,236 held-out ICU episodes from the MIMIC-III in-hospital mortality benchmark, we compared LSTM, GRU-D, XGBoost, and a weighted ensemble. The ensemble achieved an AUROC of 0.871 and AUPRC of 0.536; validation-based logistic recalibration improved the Brier score from 0.134 to 0.075. Incorrect predictions showed greater component-model disagreement and smaller decision-boundary margins, although model agreement and large margins did not guarantee correctness. Observation coverage and recent physiological trends also varied substantially across patients, highlighting differences in the information surrounding otherwise similar risk estimates. SHAP analysis attributed 89.9% of total absolute XGBoost attribution to physiological-value features and 10.1% to observation-process features. These dimensions were integrated into patient-level profiles to provide a more complete view of how predictions were formed and the clinical and data context surrounding them, extending interpretation beyond risk scores and feature rankings alone. Abbreviations: ICU (Intensive care unit)
ER -