Mitigating Sociodemographic Bias in Opioid Use Disorder Prediction: Fairness-Aware Machine Learning Framework

Area of Science:

• Computational epidemiology
• Health informatics
• Machine learning in public health

Background:

• Opioid use disorder (OUD) is a major US public health crisis affecting over 5.5 million individuals.
• Machine learning models are increasingly used for OUD risk prediction, but their fairness and potential biases are not well understood.

Purpose of the Study:

• To develop a machine learning bias mitigation algorithm specifically for sociodemographic features.
• To create a fairness-aware weighted majority voting (WMV) classifier for more equitable OUD prediction.

Main Methods:

• Utilized 2020 National Survey on Drug and Health data to train neural network (NN) models (NN-SGD and NN-Adam).
• Implemented a bias mitigation algorithm based on equality of odds to reduce disparities in model performance.
• Developed and evaluated a fairness-aware WMV classifier, alongside 1-N matching for bias analysis.

Main Results:

• The bias mitigation algorithm substantially reduced sociodemographic bias across various features (e.g., race, sex, income) for both NN-SGD and NN-Adam.
• The fairness-aware WMV classifier demonstrated high performance, achieving significant recall and accuracy rates.
• Post-matching analysis confirmed considerable bias reduction and maintained high predictive performance for OUD.

Conclusions:

• The developed bias mitigation algorithm effectively reduces sociodemographic bias in OUD prediction models.
• The fairness-aware WMV classifier proves to be a promising tool for equitable and accurate OUD risk assessment.