Interpretability and fairness evaluation of deep learning models on MIMIC-IV dataset

Area of Science:

• Artificial Intelligence in Healthcare
• Medical Informatics
• Clinical Data Science

Background:

• Large-scale healthcare datasets fuel deep learning for medical applications.
• Deep learning models raise concerns regarding interpretability, fairness, and bias in critical healthcare decisions.

Purpose of the Study:

• To analyze interpretability, dataset bias, and prediction fairness of deep learning models for in-hospital mortality prediction using the MIMIC-IV dataset.
• To connect interpretability methods with fairness metrics for quantifying disparities in mortality predictors.

Main Methods:

• Comprehensive analysis of interpretability methods for deep learning mortality prediction models.
• Evaluation of dataset representation bias and prediction fairness across demographic subgroups.
• Utilizing feature importance from interpretability to quantify fairness disparities.

Main Results:

• Interpretability methods identified key mortality predictors, including novel features beyond domain knowledge.
• Deep learning models demonstrated reliance on demographic features, indicating potential fairness issues.
• Observed disparate treatment in mechanical ventilation prescriptions and unequal reliance on racial attributes across subgroups.

Conclusions:

• Model performance alone is insufficient for healthcare applications; fairness must be prioritized.
• High prediction accuracy may stem from unfair demographic feature utilization.
• Future AI healthcare research should integrate interpretability and fairness analyses to mitigate bias.