Azra Ismail, originally from Delhi, India, was raised in Abu Dhabi, returning frequently to her roots in Bihar, one of India’s most impoverished states. Reflecting on her upbringing, she states, “I grew up witnessing stark disparities… the divide between the affluent and the underprivileged.”
Currently serving as an assistant professor of biomedical informatics at the Emory School of Medicine, Ismail is dedicated to bridging the divide between cutting-edge digital solutions like artificial intelligence (AI) and the most underserved communities in the health care sector.
A significant focus of Ismail’s research involves enhancing maternal and child health access, particularly for women from marginalized backgrounds. In partnership with the Myna Mahila Foundation and its founder, Dr. Suhani Jalota, she has developed a mobile-based AI chatbot designed to provide support for sexual and reproductive health. This includes guidance on family planning, pregnancy care, and even basic reproductive anatomy. However, adapting an AI model formulated on data from developed nations to suit a vastly different technological and cultural landscape poses its challenges.
“The main hurdles revolve around language issues and the need for these technologies to be culturally sensitive,” Ismail explains. “When health information is accessible from home in a reliable, accurate, and timely manner, it can significantly impact health outcomes. However, ensuring that this information is conveyed in a manner that is comprehensible to the end user is a demanding challenge.”
Crafting the Appropriate Model
The quest to deliver improved health care information began with extensive fieldwork aimed at understanding the specific questions and concerns of target audiences—primarily women—in the realm of sexual and reproductive health. Ismail and her team sought insights into how these concerns vary across different communities and aimed to avoid creating solutions too narrowly focused on a single demographic. “We don’t want to develop something that only serves one community,” she emphasizes.
Community health workers and user focus groups in India provided feedback that refined the questions further, leading to a pilot test conducted with 3,000 women. The next steps involve fine-tuning the language based on insights gathered from focus groups, followed by larger trials.
“It’s not merely about the data,” Ismail asserts, “but also about crafting prompts that elicit the desired responses—ensuring the tone is suitable, empathetic, and tailored to the community’s context. For instance, when discussing nutrition, it’s vital to reference locally available food sources. When suggesting social support during pregnancy, it’s essential to acknowledge prevailing gender and family norms. And when addressing contraceptive options, we must be aware of specific community challenges regarding these methods.”
Data revealed that users’ primary concerns revolved around reproduction and pregnancy, including contraceptives, family planning, safe sex practices, and health during pregnancy. Since many users communicate in Hindi but type their inquiries in English, the chatbot is designed to comprehend and respond in Hinglish, a commonly used fusion of the two languages. “To ensure uptake, it’s crucial to provide an interactive platform that allows users to engage in a familiar language format,” Ismail highlights. A resulting paper from this research has been recently accepted at the esteemed CHI conference on human-computer interaction.
Ismail credits her multicultural background with providing her insight into the obstacles that individuals face when accessing health care technologies developed in affluent countries, which may not easily adapt to diverse cultural contexts. This situation raises important questions regarding how large language models can be modified to prove effective in the developing world.
“I’m cautiously optimistic that we will witness more equitable outcomes and approaches as countries develop their own language models,” she notes. “We are observing increasing efforts from Indian researchers to create large language models that cater to languages inadequately supported by platforms like ChatGPT. Although there’s a substantial journey ahead, we are beginning to see countries take a stand on these issues and formulate AI strategies that resonate with their regional needs.”
