ISeeU

Interpretable deep learning model for ICU mortality prediction

ISeeU predicts in-hospital mortality in Intensive Care Units (ICUs) using a deep multi-scale convolutional neural network trained on the Medical Information Mart for Intensive Care III (MIMIC-III) dataset. It integrates interpretability mechanisms to explain model predictions.

Key Features:

• Deep Multi-Scale Convolutional Neural Network: Implements a multi-scale convolutional architecture for mortality prediction from MIMIC-III clinical data.
• Coalitional Game Theory-Based Interpretability: Applies concepts from coalitional game theory to generate visual explanations quantifying feature contributions to predictions.
• Validated Predictive Performance: Achieves ROC AUC of 0.8735 (±0.0025) on MIMIC-III mortality prediction tasks.

Scientific Applications:

• Clinical Outcome Modeling in Critical Care: Supports interpretable mortality risk assessment and evaluation of clinical feature relevance in ICU populations.

Methodology:

ISeeU trains a deep multi-scale convolutional neural network on MIMIC-III patient data to predict mortality and applies coalitional game theory-based attribution methods to compute feature contribution scores for model interpretability.