The Challenges of Designing Intelligent Tutoring Systems for Health Care in Low-Income Latina Populations.

Abstract

Conversational intelligent tutoring systems (CITS) have the potential to bring health education to low-income communities at scale. Although educational attainment is usually low, the health literacy spectrum within these communities can be wide. Therefore, special attention should be paid when designing and evaluating CITS within these communities. Visual, linguistic, subject matter, and cultural models must be considered. Moreover, evaluating CITS can be challenging. The instruments used to assess the impact of these systems must consider the levels of health literacy of the individuals who interact with them. In this work, we present a community-engaged design process of a breast cancer CITS for Latinas from low-income communities in Chicago. The system comprises an autonomous version and a wizard-of-Oz version-a system that appears to work autonomously to its users but is operated remotely by researchers to control for potential flaws in its artificial intelligence engine. Then, we present and contrast a case study of two women who interacted with the system. Although the two women in this paper would be traditionally grouped together for statistical analysis, the evaluations we obtained suggest that for one of them, the evaluation instruments do not seem to accurately capture her experiences interacting with the CITS. We conclude by highlighting the need for purposefully designed systems for low-income communities that rely on cultural models as well as language models and advocate for designing evaluation instruments that faithfully capture the experiences of all participants.