Fairness Correction in COVID-19 Predictive Models Using Demographic Optimization: Algorithm Development and...

Area of Science:

• Epidemiology
• Data Science
• Public Health

Background:

• COVID-19 forecasting models are crucial for resource allocation and intervention strategies.
• State-of-the-art models utilize multimodal data but suffer from under-reporting and sampling biases affecting minority groups.
• These biases lead to unfairness in COVID-19 predictions across different racial and ethnic demographics.

Purpose of the Study:

• To introduce a novel fairness correction method for aggregate-level COVID-19 case forecasting.
• To enhance the equity of predictive models used in public health decision-making.

Main Methods:

• Utilized hard and soft error parity analyses to evaluate fairness frameworks.
• Proposed and implemented Demographic Optimization (DemOpts), a debiasing method for deep learning models.
• Tested DemOpts against existing fairness correction approaches.

Main Results:

• Demonstrated significant differences in mean prediction errors across racial and ethnic groups in state-of-the-art COVID-19 models.
• Showcased that DemOpts achieves superior error parity compared to other debiasing methods.
• Confirmed that DemOpts effectively reduces disparities in mean error distributions across demographic groups.

Conclusions:

• Demographic Optimization (DemOpts) is introduced as an effective method to reduce error parity differences.
• DemOpts generates fairer COVID-19 forecasting models compared to existing literature approaches.
• The method enhances the reliability of predictions for equitable public health planning.