Explainable and Dynamic Ensemble Models for ICU Mortality and Length-of-Stay Prediction

Project Description

Predicting outcomes in Intensive Care Units (ICUs)—such as mortality risk and length of stay (LOS)—is crucial for improving care quality, reducing hospital strain, and optimizing resource use. Traditional scoring methods and classical ML approaches often struggle with the complexity, high dimensionality, and heterogeneity of ICU data.

To address these gaps, InfoLab at Sungkyunkwan University (SKKU) launched a project focused on explainable and adaptive ensemble models using multivariate time-series data from early ICU admission. The models are tailored for both adult and neonatal ICU settings and emphasize clinical relevance and transparency.

Core Contributions

1. Patient-Specific Stacking Ensemble for Adult ICU Mortality Prediction

• Developed a stacking ensemble using base learners trained on modality-specific features curated by domain experts.
• Features extracted at 6, 12, and 24 hours post-admission, enabling temporal prediction.
• MIMIC-III dataset (10,000+ patients):
  Achieved 94.4% accuracy, surpassing traditional scores (e.g., APACHE, SOFA) and standard ML baselines.
• Built a clinically interpretable pipeline integrating temporal feature slices and modality fusion.

2. Multilayer Dynamic Ensemble for Neonatal ICU Mortality and LOS

• Introduced a two-layer ensemble system:
  • Layer 1: Predicts mortality
  • Layer 2: Estimates length of stay via regression
• Employed Dynamic Ensemble Selection (DES) to adapt to the local feature space, boosting reliability under high uncertainty.
• Integrated XAI tools:
  • SHAP values
  • Decision tree visualizations
  • Rule-based logic
• Dataset: Refined NICU cohort of 3,133 neonates (MIMIC-III)
• Built a web-based clinical interface for real-time ICU monitoring and practitioner feedback

Project Objectives

• Build robust, generalizable models for early mortality and LOS prediction in adult and neonatal ICUs.
• Leverage stacking and dynamic ensemble techniques to improve accuracy and model diversity.
• Incorporate transparent explainability tools to support clinician trust.
• Tackle challenges such as missing data, irregular timestamps, and heterogeneous sources using domain-informed preprocessing.

Research Impact

By fusing clinical expertise with modern AI, this project introduces trustworthy, explainable ensemble models to the ICU:

• Enhances patient safety through early risk detection
• Improves hospital resource allocation with LOS forecasting
• Supports medical staff via actionable insights from interpretable models

Together, these efforts offer a practical foundation for next-generation critical care AI, shaping a future where real-time, data-driven decisions improve ICU outcomes and efficiency.