Using Under-Represented Subgroup Fine Tuning to Improve Fairness for Disease Prediction

Area of Science:

• Healthcare AI
• Machine Learning Ethics

Background:

• Artificial intelligence (AI) is increasingly used for disease prediction in healthcare.
• Concerns exist regarding the transparency, accountability, and fairness of AI models due to demographic disparities.
• Limited research addresses improving model fairness, especially with multivariate sensitive attributes and skewed group distributions.

Purpose of the Study:

• To explore algorithmic fairness in predicting heart disease and Alzheimer's Disease and Related Dementias (ADRD).
• To propose and evaluate a novel fine-tuning approach for enhancing fairness in predictive models.

Main Methods:

• Developed a fine-tuning approach using a pre-trained model from majority group data.
• Fine-tuned the model with data from underrepresented subgroups to incorporate specific knowledge.
• Evaluated the approach's performance against other fairness-fixing methods.

Main Results:

• The proposed fine-tuning approach outperformed existing methods across all subgroups.
• Effectiveness was demonstrated even with highly imbalanced subgroup distributions and very small subgroups.
• The method successfully incorporated subgroup-specific knowledge.

Conclusions:

• The fine-tuning approach is a promising method for improving AI model fairness in disease prediction.
• This work contributes to developing fairer AI tools for healthcare, addressing disparities in predictive modeling.
• Further research into fairness-enhancing techniques is crucial for equitable healthcare AI.