Predicting Do-Not-Resuscitate Decisions in Critically Ill Patients Through Using Multitask Learning: A Retrospective
Ming-Yen Lin1, Chuan-Feng Yeh1, Wen-Cheng Chao2,3
1Feng Chia University, Taichung, Taiwan.
Summary
Predicting do-not-resuscitate (DNR) decisions in critically ill patients is vital. An explainable multitask learning model accurately identified patients for DNR decisions, improving end-of-life care planning.
Area of Science:
- Medical Informatics
- Artificial Intelligence in Medicine
- Critical Care Medicine
Background:
- Accurate prediction of do-not-resuscitate (DNR) status is essential for ethical end-of-life care and shared decision-making in intensive care units (ICUs).
- Existing methods may lack the precision and interpretability needed for timely intervention.
Purpose of the Study:
- To develop and validate an explainable multitask learning (MTL) model for predicting DNR decisions within 24 hours for ICU patients.
- To enhance the integration of predictive models into clinical workflows to support end-of-life care discussions.
Main Methods:
- Utilized the MIMIC-IV database, training an MTL model on 7789 adult ICU patients with stays >3 days.
- Incorporated clinical parameters and nursing assessments over a 72-hour window.
- Evaluated model performance using AUROC, calibration plots, and decision curves, with interpretability via SHAP and PDP.
Main Results:
- The MTL model achieved a superior AUROC of 0.798 compared to single-task learning (0.764).
- Key predictive features included verbalization ability, ventilatory support, and muscle strength.
- Excluding a subgroup with extreme ventilatory demand and muscle weakness improved AUROC to 0.827.
Conclusions:
- An explainable MTL model can effectively predict DNR decisions in critically ill patients.
- The model's interpretability aids clinical understanding and trust.
- This approach demonstrates feasibility for integrating AI-driven insights into ICU care to facilitate DNR discussions.
