Generative AI for Health in Low & Middle Income Countries
Read the report
Case Study: Audere: 'self-care from anywhere' [Watch the launch even video here]
“Sometimes when people visit a clinic, it's really hard for clinicians to remain fully empathetic and not have some of the questions come off as abrasive, because they just don't have time to slowly gather those data. They have to ask these really personal questions … And the individuals can feel stigmatized for their answers … and just don't want to answer while they're looking somebody in the eye.” - Shawna Cooper, Director of Product, Audere
What is the GenAI use case?
Use case category: Direct-to-consumer (human-in-the-loop)
Health Areas: Communicable Diseases; Sexual and Reproductive Health
Audere is a nonprofit organization using GenAI to enhance HIV prevention and counseling services in South Africa. Their ‘Self-Care from Anywhere’ program was co-created with local community partners and SHOUT-IT-NOW, a South African nonprofit providing youth-focused HIV prevention and sexual health services. Powered by Audere’s multimodal HealthPulse AI toolkit, ‘Self-Care from Anywhere’ addresses challenges faced by adolescent girls and young women (AGYW) in accessing sexual and reproductive health services due to stigma, logistical barriers, and overstretched healthcare systems.
Accessible via WhatsApp, an empathetic AI Companion provides health education, sexual health and HIV information, self-testing guidance and care linkage, while clinicians benefit from a summarized, AI-driven decision support view of AI Companion interactions through a Clinical Portal. The AI toolkit integrates computer vision, large language models, and predictive analytics to tackle sexual health, HIV prevention, and gender-based violence. In 2025, the program will expand to include HIV treatment counseling, mental health, and TB care for additional vulnerable populations, including men living with HIV, female sex workers, and people on the edge of sex work in South Africa and Zimbabwe.
What tasks are LLMs being used for?
Summarization: Condensing client/AI Companion conversations into actionable insights for clinicians which can be viewed through the Clinical Portal.
Classification: Flagging high-risk cases for human intervention, such as disclosures of abuse or suicidal ideation.
Extraction: Retrieving key HIV vulnerability information from user conversations to inform risk assessments and tailored guidance.
Translation: Multilingual capabilities; adapting responses to local vernacular and slang to ensure cultural relevance.
Conversation: Conducting empathetic and stigma-free multi-turn interactions on sensitive topics; simulating different personas (e.g. health professional, sister, friend) according to user preference.
“Involving local partners from the beginning, co-designing with them, and ensuring local context, tone and slang and style are brought in ... We've created a slang dictionary for South Africa that has more than a thousand words in it … For instance, if a girl says, ‘I'm afraid I have the drop,’ which in South Africa means an STI, the AI Companion knows that.” - Shawna Cooper, Director of Product, Audere
Designing for inclusivity:
Community co-creation: developed in collaboration with community partners, and utilizing local guidelines to ensure context-appropriate information is provided.
Slang dictionary integration: enabling nuanced communication, trust and rapport building.
Combating stigma: focusing on reaching underserved populations with limited access to healthcare, and addressing stigma surrounding HIV.
Mitigating risks:
Conversations are monitored in real time by an automated framework, with a portion reviewed by local clinicians, HIV testing service counselors and community representatives for local relevance, accuracy, and adherence to guidelines.
High-risk cases such as disclosures of harm are flagged in real time for human intervention (clinicians).
Users are able to request to communicate with a human at any time during AI Companion conversations.
What is the current deployment status?
Early pilot, research phase:
‘Your Choice’ Study: 7-month study with 130 clients and 20 healthcare professionals, tested an early alpha version of the AI Companion for pre-HIV-self-test counseling and prevention awareness; and clinical summaries of AI/client conversations for clinicians.
‘Your Path’ Study: 12-month study with 100 clients and 25 healthcare professionals, tested a beta version for supported HIV self-testing and the AI Companion for post-test counseling on confirmatory testing or prevention options like PrEP.
Self-care from Anywhere field study: summative usability testing completed in early 2025 with 100 AGYW and 50 clinicians, with a 6-month field study commencing in Q2 2025 with 2000 AGYW in South Africa.
Across the above studies, co-design sessions, and summative usability testing, over 500 clients and nearly 100 clinicians have contributed to and used various alpha, beta, and release candidate versions of the solution.
How widely deployed could it be over time?
Target audiences: underserved populations in LMICs with limited access to healthcare and high risk and stigma surrounding HIV and SRH.
Scaling plans:
Expansion to all SHOUT-IT-NOW clients—approximately 1.5 million people in South Africa.
Expansion to all youth across South Africa through other community-based organizations and Ministry of Health support.
How are they measuring success?
LLM evaluation:
Evaluation of LLM outputs: accuracy and contextual relevance of candidate language models are assessed via an automated evaluation framework, which combines use of language models, quantitative metrics, and HITL evaluation. Once deployed, an automated monitoring framework ensures safety. •
Feasibility, acceptability, and usability for both clients and clinicians: ‘Your Choice’ trial demonstrated greater than 90% usability, acceptability and appropriateness.
Client engagement metrics (awareness, interest, intent, engagement, and retention).
User experience: qualitative data from ‘Your Choice’ trial demonstrated success in building client trust and comfort in addressing sensitive topics; healthcare providers saw value in client conversation summaries in addressing time limitations and provider mistrust.
Health outcomes:
Key health outcomes of interest—formal evaluation data forthcoming:
Improving access to education about sexual health and HIV, self-awareness of HIV status, stigma-free access to prevention or confirmatory testing options.
Improving efficiency and efficacy of clinical follow-up.
Cost-effectiveness:
Costs include usage fees for LLM services (e.g. token or API usage costs), data hosting costs, system maintenance and support, and implementation operational costs (e.g. clinician salaries). In 2023, early use of GenAI systems demonstrated between $2.65–$3.50 per 15-minute conversation, across ChatGPT and ClaudeAI versions.
System optimizations including development of an omnichannel LLM orchestration service, dynamic prompt system, and evaluation of lower cost alternatives have cut the cost to a fraction of the original token fees. Cost of use will be analyzed during the field study in 2025, and will include monitoring costs to ensure system safety, accuracy, and ethical use.
[Describing ‘Your Choice’ research study results]:
“Women were very engaged and prepared to connect with the bot, and, importantly, disclose more information than they would to human counselors. Because the disclosure was more comprehensive, the risk assessments were more accurate … In fact, many women who self-disclosed ended up realizing that their risk was significantly higher than they themselves anticipated, and opted for HIV PrEP or self-tests, or to receive HIV care. We didn't expect that kind of chain to lead to behavior change so quickly.” - Zameer Brey, Deputy Director, Technology Diffusion, Gates Foundation
Future measurement plans
‘Your Path’ study concluded in December 2024, with analysis of results ongoing.
Three scaling studies are planned for 2025, targeting different high-risk demographics for HIV counseling, prevention and treatment support. These also include expanded language support for isiZulu and Shona, and will include evaluations of commercial vs. fine-tuned LMs for each task where LLMs are utilized.
A/B testing between intervention arms with and without GenAI integration.
Evaluation of GenAI computer vision capabilities vs. purpose-built models for rapid test identification and interpretation via the automated evaluation framework.
Plans to evaluate savings of preventive intervention against normal treatment costs.
See report to read the full article including other case studies
