Health equity assessment of machine learning performance (HEAL): a framework and dermatology AI model case study

Area of Science:

• Health AI and Machine Learning
• Medical Informatics
• Health Equity Research

Background:

• Artificial intelligence (AI) in healthcare can perpetuate historical inequities.
• Existing fairness metrics do not fully capture performance equity for diverse patient groups.

Purpose of the Study:

• To develop a quantitative framework, the Health Equity Assessment of machine Learning performance (HEAL), to assess the performance equity of health AI.
• To evaluate the utility of the HEAL framework using a case study of a dermatology AI model.

Main Methods:

• The HEAL framework employs a four-step interdisciplinary process to define and quantify domain-specific criteria for performance equity.
• A case study applied HEAL to a dermatology AI model using 5420 teledermatology cases, assessing performance across diverse age, sex, and race/ethnicity subpopulations.
• The HEAL metric was calculated as the probability that AI performance is better for subpopulations with worse average health outcomes, using DALYs and YLLs to quantify health disparities.

Main Results:

• The dermatology AI model demonstrated a HEAL metric of 80.5% for race/ethnicity, 92.1% for sex, and 0.0% for age, indicating prioritized performance for certain subpopulations.
• For skin cancer conditions, the HEAL metric was 73.8% for prioritizing age subpopulations.
• The AI model's performance was associated with pre-existing health disparities across race/ethnicity, sex, and age (for cancer conditions).

Conclusions:

• The HEAL framework quantitatively demonstrated that the evaluated dermatology AI model prioritized performance for subpopulations facing health disparities.
• Further research is needed to promote equitable AI performance across all age groups and for non-cancer conditions.
• Understanding how AI can improve health equity requires continued investigation into AI model contributions.