Evaluation and improvement of algorithmic fairness for COVID-19 severity classification using Explainable Artificial...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Health Equity

Background:

• Machine learning (ML) models are increasingly used for predicting COVID-19 severity, crucial for clinical decisions and resource allocation.
• Ensuring fairness in ML predictions is vital to prevent healthcare disparities, especially concerning sex-based biases.
• Existing fairness interventions often reduce model accuracy, limiting clinical applicability.

Purpose of the Study:

• To evaluate fairness in an ML-based COVID-19 severity classification model.
• To propose and assess an Explainable AI (XAI)-based strategy for mitigating sex-related bias.
• To achieve a balance between predictive accuracy and fairness in clinical decision support systems.

Main Methods:

• Developed an XGBoost multi-class classification model using Quebec Biobank data.
• Assessed fairness using Subset Accuracy Parity Difference (SAPD) and Label-wise Equal Opportunity Difference (LEOD) metrics.
• Implemented and compared four bias mitigation strategies, including an XAI-based method utilizing SHapley Additive exPlanations (SHAP).

Main Results:

• The baseline model achieved 90.68% accuracy but showed a 10.11% Subset Accuracy Parity Difference between sexes.
• The XAI-based method demonstrated a superior trade-off between performance and fairness compared to other strategies.
• Identified and integrated sex-sensitive feature interactions into model retraining using SHAP values.

Conclusions:

• The XAI-driven bias mitigation effectively reduces sex-based disparities in COVID-19 severity prediction.
• This approach minimizes accuracy loss compared to traditional fairness interventions.
• Provides a framework for developing fair and accurate clinical decision support systems for equitable care in older adults.