Google researchers have introduced a novel framework called HEAL (Health Equity Assessment Likelihood) to evaluate how artificial intelligence (AI) tools impact health disparities across different demographic groups. The framework aims to quantify the likelihood that an AI model prioritizes performance for subpopulations with worse health outcomes.
In a study published by Google, the HEAL framework was applied to a dermatological AI model, revealing that while the model performs well across sex and race/ethnicity groups, it has room for improvement when it comes to age groups, particularly for non-cancer skin conditions. For instance, individuals aged 70 and above, who face the poorest health outcomes related to non-cancer skin conditions, were not prioritized in terms of model performance.
The HEAL framework evaluates AI models based on pre-existing health disparities and model performance metrics, such as the top-3 agreement rate. It was found that the model was 80.5% likely to perform equitably across race/ethnicity subgroups and 92.1% likely to perform equitably across sexes. However, the model showed less equitable performance for age groups in non-cancer conditions.
The framework, however, is not meant to be used in isolation. Researchers emphasize that it should be contextualized with other factors such as computational efficiency, data privacy, ethical values, and potential selection biases. An adversarial example was presented to illustrate that the HEAL metric can be artificially improved by reducing performance for the most advantaged group, which could lead to overall worse performance.
To address this, the HEAL framework should be used alongside a Pareto condition, ensuring that no subpopulation's outcome worsens compared to the status quo. The researchers acknowledge that while HEAL can assess the likelihood of AI models prioritizing performance for disadvantaged groups, it cannot determine whether AI will actually reduce disparities in outcomes in real-world settings. This requires a causal understanding of the entire care journey.
The HEAL framework is proposed as a valuable tool not only during model development but also during pre-implementation and real-world monitoring stages. It could potentially be implemented in the form of health equity dashboards to track and improve AI models' performance across different populations. The framework's strength lies in its adaptability and potential for refinement through future applications.
Google researchers note that a successful approach to understanding AI's impact on health equity requires more than just metrics. It necessitates a community-driven set of goals that represent those most impacted by the models. The HEAL framework is seen as a significant step toward addressing the grand challenge of AI and health equity, offering a structured method to evaluate and improve AI technologies for better health outcomes across all demographics.
Source: Google Research Blog
