The primary objective of health equity is to ensure that every individual has an equal and just opportunity to achieve their optimal health. Unfortunately, many individuals, including people of color, women, those in rural areas, and other historically marginalized groups, do not experience this equity. As the Chief Health Equity Officer at Google, my team is dedicated to ensuring that we develop AI-powered health solutions responsibly and equitably.
During our annual health event, The Check Up, we introduced three approaches that aim to usher in a more equitable future.
Recent research on identifying and addressing biases
In light of the rapid evolution of medical AI, it’s crucial to create tools and resources that can detect and rectify biases that might adversely affect health outcomes. Our latest research paper, titled “A Toolbox for Detecting Health Equity Issues and Biases in Large Language Models,” represents a significant step in this direction. This paper presents a framework for assessing whether medical large language models (LLMs) may perpetuate historical biases and offers a set of seven testing datasets known as “EquityMedQA” as a guideline.
These tools are grounded in literature on health disparities, documented model deficiencies, and insights from equity experts. We utilized these tools to evaluate our own large language models, making them accessible to the broader research community.
Introduction of a new framework for evaluating health equity in AI models
A diverse group of experts from various fields, including health equity researchers, social scientists, clinicians, bioethicists, statisticians, and AI specialists at Google, collaborated to develop a framework aimed at constructing AI systems that avoid creating or perpetuating unfair biases.
Known as HEAL (Health Equity Assessment of Machine Learning performance), this framework assesses the likelihood of AI technologies performing equitably and prevents the deployment of AI models that could exacerbate disparities, particularly for groups with historically poorer health outcomes. The process comprises four key steps:
- Identifying factors linked to health inequities and establishing AI performance indicators.
- Quantifying existing health outcome gaps.
- Assessing AI tool performance for each subpopulation.
- Evaluating the extent to which the AI tool prioritizes performance concerning health disparities.
We have already applied this framework to evaluate a dermatology AI model. The outcomes revealed equitable performance across race, ethnicity, and gender subgroups. However, enhancements were identified to better serve older age groups. The framework showed that while the model demonstrated fair performance in diagnosing cancerous conditions such as melanoma across age groups, its performance was not as strong for non-cancerous conditions like eczema in individuals aged 70 and above.
Moving forward, we will continue to use this framework to assess healthcare AI models, refining and enhancing it along the way.
Enhanced dermatology dataset for comprehensive analysis
Currently, many dermatology datasets lack diversity, hindering the development of fair AI models. Existing dataset images are often obtained in clinical settings and may not represent various body parts, different condition severities, or diverse skin tones, ages, and genders. Moreover, they predominantly focus on severe conditions like skin cancer, neglecting common issues such as allergic, inflammatory, or infectious conditions.
To create a more inclusive image dataset, we collaborated with Stanford Medicine to establish the Skin Condition Image Network (SCIN). Over 10,000 real-world dermatology images were contributed by thousands of individuals to form this openly accessible dataset. Dermatologists and researchers then annotated diagnoses and categorized them based on two skin-tone scales, ensuring an extensive range of conditions and skin types were included.
The SCIN dataset can now be utilized by scientists and medical professionals to aid in the identification of dermatological issues, conduct dermatology-related research, and expose healthcare students to diverse examples of skin conditions across various skin types.
While we are in the early stages of this initiative, we are committed to effecting positive change. By collaborating with partners and sharing our insights, we strive to contribute to a healthier future for all individuals, irrespective of their backgrounds or geographic locations.
