From Clinic to Code: Using Clinician Insights to Develop a Framework for Fair and Representative Datasets in Women’s Health AI

The development of Artificial Intelligence (AI) in healthcare is largely dependent on the quality of medical datasets. However, these datasets often fail to accurately represent women (and those seen as women) due to historic, implicit and biological biases. This under-representation can lead to biased, inequitable and even harmful models. Building upon the findings of previously conducted qualitative semi-structured semantic interviews with clinicians on their perceptions of women’s health, this paper presents a framework for translating qualitative findings to dataset characteristics via operationalisation. This framework outlines the key characteristics and considerations a dataset should include or consider to more accurately represent women in these datasets. Some of these factors include: pregnancy status, gender of provider of the care, menstruation and menopause, and ethnicity. Rather than considering fairness after model development and employing de-biasing metrics, this approach places fairness at the initial selection stages, with the goal of embedding equity throughout the entire development pipeline. It is both a checklist for data selection for model developers and also a guideline for those who collect medical data. The framework is divided into Necessities, Data Comprehension, Gender-Specific Factors, Clinician Information, Patient-Specific Factors, and Additional considerations. This framework is a step towards creating more gender-conscious, equitable and fair medical AI systems.