Awards

Michael Leventhal of the Association RobotsMali in Mali will determine whether ChatGPT-4 can support curriculum development and teacher training to improve literacy in Mali, which has 65% illiteracy. The West African language Bambara is the most widely spoken language of Mali, but there is almost no literature in Bambara and few Malians can read their mother tongue. Education is provided almost entirely in French, a language most Malians do not understand, and in a cultural context foreign to Malian children. They will use ChatGPT-4 to generate graded, culture-specific written stories for children in Bambara along with linked pedagogical material for teachers to improve lesson quality. They will evaluate the material with students and teachers using available quantitative tools and assess its ability to improve educational outcomes.

Neal Lesh of Dimagi South Africa (Pty) Ltd. in South Africa will create an LLM-powered coach tailored to frontline workers that offers training, performance feedback, and encouragement to support their health and improve their productivity. Frontline programs serve billions of people; however, they rely on a hard-working, often overburdened workforce that receives limited support, particularly in low- and middle-income countries. They will work with 10–20 community health volunteers in Malawi to co-design three variations of the LLM-powered coach using their rapid LLM-building platform. They will assemble content on early childhood development and the Kangaroo Mother Care method. They will then design professional development curriculums to strengthen existing skills; teach new skills, such as financial management; and build resilience skills to encourage self-care and well-being. They will test the coaching bots on 100 frontline workers to evaluate safety, accuracy, usability, and added value.

Bishesh Khanal of the Nepal Applied Mathematics and Informatics Institute for Research in Nepal will assess LLMs for their ability to provide accurate information on sexual, reproductive, and maternal health (SRMH) topics in Nepali to the general public and female community health volunteers. In Nepal, limited access to SRMH resources due to language barriers and social stigmas has led to increased numbers of unsafe pregnancies and sexually transmitted diseases. While LLMs could be helpful, they have many limitations, particularly in low-resource, non-Western settings. These include inaccurate responses, poor performance in non-English languages, responses generated largely from Western-cultural contexts, and large computational resource requirements. Together with a local multidisciplinary team, involving AI scientists, domain experts, and community engagement experts, they will integrate four chatbots into a simple mobile-friendly web-interface, and evaluate their performance to anonymous chat queries from 5,000 individuals.

Leonora Tima of Kwanele - Bringing Women Justice in South Africa will develop a mobile application and chatbot to provide understandable legal information on gender-based violence (GBV) to vulnerable groups, including high school learners, young women, survivors of GBV, members of the LGBTQIA+ community and sex workers. South Africa faces disproportionately high rates of GBV but lacks access to justice and understandable legal information for survivors. They will integrate GPT4 and OpenAI's Large Language Model (LLM) with front-end applications, such as WhatsApp and Facebook, to guide users through the complex judicial system using everyday language. They will run community training and onboarding events to demonstrate the technology and introduce people to the application and they will run focus groups, workshops and interviews to support the design of the tool and build the datasets.

Scott Mahoney of The Health Foundation of South Africa will create an application that combines human expertise with AI technology to produce clinical recommendations from published medical evidence to be used as a decision-support tool for healthcare professionals in low- and middle-income countries. Currently, producing guidelines and support tools relies on manual reading and synthesis by individual clinicians or editorial teams, which is time consuming and can lead to biased coverage. The application will use ChatGPT-4 and be able to analyze text-based medical evidence in various formats, extract relevant clinical recommendations, and formulate clinician-validated clinical decision support algorithms for frontline healthcare workers in near-real time. This will improve the speed, accuracy, and inclusivity of decision-making. They will validate the application by comparing its performance with recommendations provided by their clinical editorial team.

Suhani Jalota of the Myna Mahila Foundation in India will build a chatbot, Myna Bolo, by incorporating Large Language Models (LLMs) into their health application to provide tailored sexual and reproductive health services through smartphones, via text or audio, in local languages to women in India. In India, 71% of girls report not knowing about menstruation before their first period. This is because of limited access to unbiased information due to stigma, discrimination, and lack of resources. Information needs to be non-judgmental, confidential, accurate, and tailored to those living in urban slums. They will incorporate LLMs by integrating Google Bard into their application. Women can then ask questions and receive tailored responses that are considerate of their backgrounds and limited smartphone access, and respectful of their privacy. They will select accurate source material, relevant to local women in different languages, and incorporate fact-checking capabilities and maps for providing referrals and treatments.

Henrique Dias of the Instituto de Inteligencia Artificial na Saude in Brazil will determine whether AI can produce an accurate hospital discharge summary to ensure that essential information is passed to the next healthcare provider and patient care is maintained. Discharge summaries are often incomplete, unclear, or delayed in terms of their delivery due to the document construction process. They will test two AI models to produce discharge summaries - one trained by health professionals completing electronic medical records, and the other trained with 46 GB of data in Portuguese, corresponding to 38 million clinical notes from 70 hospitals. They will perform a retrospective, non-inferiority, single-blind study to compare the quality and speed of the discharge summaries produced by both AI models with those produced by medical professionals.

Mamadou Alpha Diallo of Cheikh Anta Diop University in Senegal will apply Large Language Models (LLMs) to improve decision-making, policy development, resource allocation and communication to help combat infectious diseases in Africa. They will use ChatGPT-4 to analyze and interpret epidemiological data, clinical records, and research literature to help predict outbreaks, identify priority areas for interventions, and evaluate the potential impacts of specific policies. The information produced will include tailored messages, educational materials, and real-time updates on disease trends and prevention strategies for healthcare workers, policymakers, and affected communities. This tool is expected to achieve the following impact in LMICs: enable faster, more accurate, and more inclusive decision making; strengthen the healthcare system at all levels from the healthcare worker to the policymaker; result in the reduction of disease burden; and reduce healthcare disparities through enhanced equity and access to information, resources, and interventions.

Daudi Jjingo of the Infectious Diseases Institute in Uganda will leverage generative AI to develop an interactive conversation-based platform to communicate the national guidelines for pandemic preparedness in a native African language to health workers to improve pandemic management. The national guidelines, currently available as a lengthy PDF, will be translated into a local Bantu language, Luganda, to improve accessibility to non-English speaking users, and converted into a data format for Large Language Models (LLMs) such as GPT-4. The data will include locally-relevant, medically-curated, and pre-approved information for pandemic preparedness, including prevention, detection, and treatment strategies. They will also build a user-friendly, dynamic interface for health workers to interact with the AI model as needed and use the information to guide interventions. Their platform will be field-tested by a group of 60 health workers at 20 clinical sites.

Joyce Nakatumba-Nabende of Makerere University in Uganda will leverage ChatGPT to provide tailored support to smallholder farmers in sub-Saharan Africa in their local languages. These smallholder farmers contribute up to 69% of household incomes, but they are vulnerable to the devastating effects of crop diseases and pests and lack the timely support required to combat such challenges. Digital technologies have been developed to help but they cover a limited number of crop types and languages. They will curate a dataset comprising 1,000 farmer-specific agricultural questions on pests and diseases, markets, and seed advisory services, in English and Luganda. They will then perform prompt engineering of ChatGPT to investigate its potential to provide targeted, accurate, and unbiased agricultural advice on a wide range of crops. The responses will be further fine-tuned and compared with responses from agricultural experts.

Cally Ardington of the University of Cape Town in South Africa will develop an AI-powered voice-recognition model that performs Early Grade Reading Assessments (EGRA) in low- and middle-income countries (LMICs). Seventy percent of children in LMICs do not learn to read in any language, which severely affects their overall education and future prospects. Reading assessments, such as EGRA, test children on letter-sound knowledge, word reading, reading connected text, and answering questions on that text. They are critical for supporting reading programs but are currently expensive and time-consuming because they are administered one-on-one. They will perform a pilot study to determine whether a new open-source voice recognition program developed by Facebook (wav2vec), which is especially useful for languages with little training data, can automatically evaluate speech production and assess children's early reading abilities in African languages. They will validate EGRA-AI by adding it to an existing 120-school field trial using standard EGRA.

Hugo Morales of Munai Health in Brazil will integrate OpenAI's ChatGPT-4 and other Large Language Models (LLMs) with Munai's Clinical Intelligence platform to help frontline healthcare providers adhere to guidelines for antimicrobial therapy and reduce antimicrobial resistance. Antibiotic-resistant bacteria cause over 20% of infections in Brazil. However, the antimicrobial stewardship programs designed to address this consist of complex protocols and there is little training for health workers in low-resource settings. The Munai platform incorporates machine learning into an application and web interface that is currently connected to 29 hospitals and hosts data from over 12 million medical encounters. They will incorporate conversational AI tools into the platform to enhance accessibility and usability. They will then convert institutional antimicrobial therapy protocols into a machine-readable format and incorporate patient data to allow for more personalized responses. The LLM will be evaluated by a prospective study engaging 20 physicians in a two-part simulated clinical trial.

Enrica Duncan of Mapa Do Acolhimento in Brazil will use AI to improve the influx of volunteer psychologists and lawyers to their support network, which provides mental health and legal support to women at risk of gender-based violence. In 2022, one woman died every six hours from gender-based violence in Brazil. They have built a network of 10,000 volunteers who have supported over 5,000 women. To expand this network, they have developed core training content using an inter-sectional feminist framework and will produce an equitable AI model to evaluate volunteers' backgrounds to better understand their skills and experiences. They will also incorporate localized knowledge to better serve all regions of Brazil. Machine learning will also be used to determine the optimal format to deliver training for each individual based on their engagement and responsiveness. These efforts aim to enhance the volunteer experience, improve knowledge absorption, and reduce turnover.

Tonee Ndungu of Kytabu Company Ltd. in Kenya will develop a comprehensive AI-powered mobile application, SOMANASI (derived from the Swahili words meaning "learn together") to provide personalized education to every student in Kenya. Kenya suffers from widespread educational inequities with many students failing to receive individualized attention. The application will harness ChatGPT-4 and act as an intelligent virtual tutor that delivers tailored content, adaptive learning experiences, and interactive guidance. They will collaborate with experts to design high-quality materials aligned with the Kenyan curriculum and cultural context. They will also engage students, teachers, and educational stakeholders in the design process, and mitigate bias by considering the full diversity of the student population. They will pilot test SOMANASI across a diverse student population in ten schools to evaluate its ability to enhance learning outcomes.

Moinul Haque Chowdhury of CMED Health Limited in Bangladesh will integrate a multilingual AI engine into their existing digital healthcare platform, SuSastho, to produce a chatbot that provides secure access to sexual, reproductive, and mental health care for adolescents. Bangladesh has the highest adolescent pregnancy rates globally, and 16-18% of its adolescents suffer from mental disorders; however, little to no sexual, reproductive, or mental health care is available. They will use an open-source language model that operates in multiple languages, including Bangla. Together with experts, they will compile common queries and sought-after information regarding education, early marriages, contraceptive use, adolescent pregnancies, and sexual and mental health, and collect training data for the AI model. The chatbot will also be designed to assess health risks and make referrals. They will conduct beta testing, clinical validation, user acceptability testing, and cultural validation through consultative workshops.

Suzanne Staples of the THINK Tuberculosis and HIV Investigative Network (RF) NPC in South Africa and Kristina Wallengren of THINK International in Denmark will produce a toolkit that leverages ChatGPT for the analysis and interpretation of health program data in low- and middle-income countries (LMICs). Due to resource constraints, data analysis takes a back seat to diagnostics and treatments and is a scarce skill in LMICs, particularly in the public health sector. In addition, health data management is hindered by manual and fragmented electronic datasets. They will work with end-users, including program managers and decision-makers, to generate a toolkit that utilizes ChatGPT in data analysis to drive evidence-based decision-making, and aid in the early detection of disease outbreaks, initially focusing on the TB program. They will assemble the most frequent queries by various stakeholders to identify real priorities for program management and evaluate the ability of ChatGPT to analyze and interpret routine TB program data.

Christophe Bocquet of Dalberg Global Development Advisors (K) Ltd. in Kenya will develop VIDA PLUS, a chatbot accessible via WhatsApp that delivers public health information by live interaction to health officials, particularly in rural areas, to support their decision-making. Accessing relevant public health information is often challenging for health workers in rural areas who have limited access to technology and data literacy. Initially in Guinea, they will integrate GPT-3.5-Turbo into the national health management information system (HMIS), which comprises data on health outcomes, health facilities and utilization, and disease surveillance. This will enable health officials to ask questions on topics such as maternal health, infections, vaccinations, and hospitalization, and receive tailored answers via WhatsApp. Health officials will be involved in the design, deployment, and testing stages, and they will also plan the scale-up, including a cost and impact analysis.

Robert Korom of Penda Health Limited in Kenya will integrate ChatGPT-4 into their established patient communication system to increase consultation efficiency and the speed of delivering accurate health information in Kenya. Their existing chat-based digital health solution relies on a dedicated team of clinicians and call center agents to serve low-income Kenyans; however, increasing needs are leading to longer response times. They propose to blend the empathetic and intuitive nature of human interaction with the instantaneous, data-driven capabilities of AI to improve throughput, response times, and patient experience. This will create a hybrid model where clinical call center agents work hand-in-hand with AI. They will carry out a proof-of-concept involving a limited field test to monitor patient satisfaction, efficiency, and relevance of responses.

Olubayo Adekanmbi of Data Science Nigeria in Nigeria will develop a multilingual, voice-based chatbot to demystify complex financial concepts and provide customized financial support to informal traders, women business owners, and smallholder farmers in Nigeria. These groups are often disadvantaged due to their low income and literacy and are historically underserved by conventional financial systems. They will create a chatbot capable of recording transactions from verbal inputs, such as "I bought four oranges at N50 naira," and answering financial questions. Customized financial guidance will be communicated back to the end user via voice note using text-to-speech. They will engage users in the design process and build large, locally-orientated financial datasets. They will then merge speech-to-text technology and GPT learning capabilities with an AI-driven financial management tool. The chatbot will be tested using their existing network of informal traders, and the feedback used for refinement and improvement.

Joseph Mulabbi of Comzine Tech And Investments Limited - Dromedic Health Care in Uganda will use ChatGPT-4 to optimize the surveillance of zoonotic diseases and predict future pandemics. Zoonoses are infectious human diseases that originate from animals and represent over 75% of all emerging diseases. Predicting the emergence of a zoonotic disease currently requires manual monitoring of the dynamic interactions between humans and livestock, which is time-consuming, resource-intensive, and prone to delays. Together with relevant stakeholders, they will build DROMEDIC-AI, a ChatGPT-4 AI platform trained with large volumes of text from diverse sources, such as news articles, social media, and clinical notes, where farmers can upload photos of sick animals and receive advice. The platform will also generate risk assessments and maps of hotspots and inform health officials to help them better monitor potential outbreaks. They will collect user interaction data and feedback on its performance.

Chinazo Anebelundu of DSN Ai Innovations Limited in Nigeria will develop a Science, Technology, Engineering, and Mathematics (STEM)-focused multimedia learning platform by leveraging GPT and DeepBrain text-to-video AI tailored to rural students to increase their engagement. Nigeria has an estimated 18.5 million students out of school, a population that can potentially be reached through this more engaging and personalized modality. This platform will integrate local contexts and nuances to enhance student comprehension of STEM subjects. It will also tailor the learning to the preferences of each student and employ visual activities such as interactive STEM laboratory simulations. For each subject, they will collect a diverse array of educational resources and construct the curriculum. This curriculum will then be transformed by Large Language Models (LLMs) into a script and then into video lessons that can be personalized to a specific student. The platform will be evaluated for the degree of student engagement and accuracy of content, amongst other measures, using their existing network of students and teachers.

Brenda Hendry of Boresha Live in Tanzania will integrate ChatGPT-4 into community radio to broadcast inclusive health messages across Tanzania to combat malaria. Tanzania is among the top ten countries with the highest malaria cases and deaths. Their control efforts are severely hampered by limited access to accurate health information among certain populations. Radio is very popular and reaches across rural and remote areas making it a powerful communication medium. To leverage this, they will train ChatGPT-4 with malaria health information and local contexts and behaviors collected from community leaders, healthcare professionals, and malaria control programs. ChatGPT-4 will then be able to produce accurate malaria-related information that respects cultural norms, language preferences, and local challenges. These messages will be broadcast by community radio stations that reach over 36 million people, and they will evaluate the impact on increasing knowledge and reducing malaria cases.

Amelia Taylor of Malawi University of Business and Applied Sciences in Malawi will employ Large Language Models (LLMs), including ChatGPT and MedPalm, to develop a tool to streamline the collection, analysis, and use of COVID-19 data. Collecting accurate and comprehensive data during a pandemic is critical for response efforts but the process is labor-intensive. During COVID-19 surveillance, there were also limited training materials available to explain specialized concepts for data collection to the multidisciplinary teams. To address this, they will leverage their experience in the operational aspects of COVID-19 data management in Malawi to develop an automated feedback tool that records high-quality, complete data while ensuring compatibility across diverse sources and destinations. Furthermore, together with frontline health workers and epidemiologists, they will create a knowledge map of symptoms and clinical terminologies to support clinicians, lab technicians and surveillance officers engaged in data collection.

Darlington Akogo of MinoHealth AI Labs in Ghana will leverage a multimodal Large Language Model (LLM) to generate accurate and comprehensive medical reports based on the analysis of medical images to reduce the need for manual reports and enhance diagnostic capabilities for radiologists and clinicians. African healthcare systems have excessively high patient-to-doctor ratios and prevalent diseases and severely inadequate numbers of radiologists. They will fine-tune a multimodal LLM applied to radiology and medical imaging using a supervised approach with a labeled dataset of medical images and corresponding reports collected from facilities across Ghana and Africa. The platform will enable interactive conversations with clinicians seeking answers to specific queries or clarifications regarding medical images. They will use metrics and humans to evaluate the model and assess its ability to generate accurate and comprehensive medical reports. They will also conduct field testing with clinicians and individuals from diverse demographics.

Essa Mohamedali and Kalebu Gwalugano of the Tanzania AI Community in Tanzania will use ChatGPT-4 to develop a chatbot and support tool to help healthcare workers adhere to the Integrated Management of Child Illness (IMCI) guidelines and access updates and alternative treatment options by linking them to the latest research via their mobile phones. Access to formal training on the IMCI guidelines is limited for healthcare workers, particularly in the private sector, and its duration makes it prohibitively expensive for companies. They will convert the existing guidelines and algorithms into a chatbot version and use the GPT-4 framework to connect to the latest research. They will engage healthcare workers during the development stage and then implement and field test the support tool at three private health facilities in rural, urban, and peri-urban areas of Tanzania, to assess its usability.

Amrita Mahale of ARMMAN in India, in collaboration with colleagues at ARTPARK also in India, will integrate an LLM-powered co-pilot into an existing learning and support application to improve the training of auxiliary nurses and midwives in India so they can better manage high-risk pregnancies. One woman dies in childbirth every twenty minutes in India. Many maternal and infant deaths could be prevented by improving access to critical care information and ensuring that health workers can detect risk factors and treat complications early on. They will fine-tune ChatGPT, or an open-source equivalent, with their existing content in English and Telugu, using machine translation models to provide personalized answers depending on the individual's training level. The design will be informed by user research studies and the application will be trialed with over 100 community health workers.

Minh Do of Fulbright University Vietnam in Vietnam will create a chatbot "NướcGPT" (Nước means water in Vietnamese) that combines cutting-edge AI tools with a user-friendly interface in the local language to support the management of salinity intrusion in the Mekong Delta. The Mekong Delta, home to 21.5 million Vietnamese, is suffering from increased saltwater intrusion caused by multiple factors including climate change. They will fine-tune GPT3.5 and GPT4 using data in English and Vietnamese collected from diverse sources including literature reviews, field studies of water-related problems, and technological solutions. This will bridge the gap between complex scientific knowledge and practical decision-making, empowering users to make informed choices. They will also build a website to host their chatbot and field test it with selected stakeholders, including government officials and farmers to evaluate its usability, accuracy, and ability to support decision-making processes.

Floris Sonnemans of Degas Ghana Limited in Ghana will apply AI technology to support African smallholder farmers to implement more climate-adaptive and regenerative agricultural (RA) techniques, such as crop diversification, and scale climate action across the continent. Africa only contributes 3.8% of global greenhouse gas emissions but experiences the harshest impacts, particularly on food production. However, protective RA techniques are relatively new and challenging to adopt, and there are not enough field agents to support farmers and respond to queries. To address this, they will integrate a Large Language Model (LLM)-based system trained on RA manuals into their existing agent-facing application. This system will provide an easy interface for farmers to access information and guidance to effectively implement RA practices, such as biochar application, minimal tillage, and permanent organic soil coverage. The application will be field tested amongst field agents and farmers.

Nana Kofi Quakyi of the Aurum Institute in Ghana will develop an AI-powered decision support tool for antibiotic prescribers to improve appropriate antimicrobial usage and combat antimicrobial resistance (AMR) in Ghana. AMR is a major public health concern, with the highest mortality rates occurring in Africa. To address this, Ghana's National Action Plan for Antimicrobial Use and Resistance (2018) identified the need for support tools that provide personalized, adaptable, and context-sensitive recommendations. They will develop an interactive, AI-powered clinical decision support tool that allows prescribers to enter prompts, respond to system queries, and receive personalized, real-time antibiotic prescribing recommendations, such as drug, dosage, and duration of administration. The model will be trained with a comprehensive dataset consisting of clinical guidelines, research data, expert opinions, and surveillance evidence that has been reviewed for its representativeness. The proposal includes field testing, iterative refinement, and engagement with stakeholders to ensure effectiveness and scalability.

Nirat Bhatnagar of the Belongg Community Ventures Private Ltd. in India, in collaboration with colleagues at ARTPARK also in India, will develop a Large Language Model (LLM)-based tool to enable development practitioners, funders, and researchers to adopt more equitable approaches, particularly addressing the intersections of marginalization. They will assemble a comprehensive and trusted corpus of development research papers, reports, and media articles and use it to build a user-friendly website and a backend ChatGPT 4.0 API-based LLM model. Users will be able to upload their draft research or program proposals and receive tailored recommendations on how to increase inclusivity across dimensions such as gender, disability, caste, religion, ethnicity, and sexual orientation. The tool will also connect users with researchers and experts with experience living with marginalized identities.

Yogesh Hooda of the Child Health Research Foundation in Bangladesh will use AI-based tools to teach low- and middle-income scientists to perform modeling and prediction studies in public health, which are dominated by researchers in the Global North. The codes generated during modeling studies are not often shared amongst researchers, making the methods difficult to learn. They found that ChatGPT could produce a code using a published model in just three weeks with only a beginner-level programmer and a biostatistician. Using epidemiological and demographic data and medical records collected from a catchment area, they will adapt published code with the help of ChatGPT to predict the impact of introducing specific vaccines in Bangladesh. They will also develop a curriculum, covering the basics of ChatGPT, data preprocessing and modeling techniques, for a course that they will pilot with public health professionals and students. All materials will be openly available in Bangla and English.

Faisal Sultan and Sara Khalid of Shaukat Khanum Memorial Cancer Hospital and Research Centre in Pakistan will leverage the power of open-source AI Large Language Models (LLMs) to extract insights more quickly and easily from large volumes of clinical data to support medical decision-making and minimize health disparities in South Asia. Healthcare systems in South Asia have limited resources and the critical information required for decision-making is often buried in patient notes (such as family history, drug adverse events, and social, behavioral, and environmental determinants). Health disparities are also prevalent, particularly for women and children. They will leverage existing LLMs specifically designed for health data and use the SKMCH&RC database, which contains both free text and structured data for 250,000 patients, to ensure that key subjective information, such as family history, and electronic health records are included. They will validate their model using available data on COVID-19 infections in Pakistan and evaluate its performance in terms of accuracy and speed.

Khoa Doan of VinUniversity in Vietnam together with Helen Meng, Viet Anh Nguyen, and colleagues from the Centre for Perceptual and Interactive Intelligence and The Chinese University of Hong Kong, both in Hong Kong; in collaboration with the Hanoi Obstetrics & Gynecology Hospital in Vietnam, will build a conversational AI chatbot to scale up gynecological healthcare support for women and LGBT+ communities in Vietnam. Access to gynecological healthcare in Vietnam is limited, particularly in remote regions and for minority groups due to high costs, low investment, social stigmas, and misinformation. They will build a tool to provide informational and psychological support, adapted to Vietnamese linguistic and cultural contexts and able to operate in low-resource settings. The chatbot will consist of a scientific database, a GPT-like conversational system, a voice generation engine, and a sentiment analytics module to evaluate the psychological traits of the user. It will be capable of empathetic dialogues that encourage users to share their symptoms. Data from patient-clinician dialogs will be used as a reference for the design of an initial patient-AI prototype.

Nneka Mobisson of mDoc Healthcare in Nigeria will integrate ChatGPT-4 into their chatbot, Kem, which provides virtual self-care coaching for low-income women of reproductive age in Nigeria, to improve its accuracy and capacity to respond to queries with evidence-based information. The burden of maternal deaths in Nigeria remains inequitably high with many risks encountered even before conception, highlighting the importance of supporting self-care. They will assess the accuracy and empathy of Kem+ChatGPT's responses to diverse inquiries from women, its improved ability to triage based on reproductive stages and risk factors, and the effectiveness of human health coaches leveraging ChatGPT-4 as a resource for answering more complex questions. This will involve field testing with a cohort of 300 women of reproductive age. They will also hold a series of iterative user-testing workshops across the states with at least 160 members from communities to observe usability and provide feedback for product refinement.

Martin Mwangi of Intellisoft Consulting Ltd. in Kenya will build an application-supported LLM to improve knowledge, attitudes, and practices surrounding the risk factors for non-communicable diseases (NCD) for young people in Kenya. NCDs constitute the leading cause of mortality globally, accounting for three-quarters of deaths worldwide. Many Kenyans lack information on NCDs and their major risk factors, which include unhealthy diet, physical inactivity, and harmful alcohol use. They will form an interdisciplinary Community Advisory Board, including government officials, researchers, and young people, to guide the design, analysis, and dissemination of the app. They will recruit Kenyans aged 18–34 from community-based sites, such as universities and malls, to evaluate the application's ability to improve knowledge, attitudes, and practices surrounding NCD risk factors.

Daphne Ngunjiri of Access Afya in Kenya will integrate ChatGPT into a virtual clinic application, mDaktari, to support clinicians and better respond to patient inquiries. Poor quality healthcare results in 5.7 million deaths in low- and middle-income countries, emphasizing the need to increase healthcare quality as well as accessibility. Their mDaktari platform combines a digital and physical healthcare network, telemedicine, and localized patient health data to support patients and clinicians in low-income communities from diagnosis to treatment. They propose to scale their approach using Large Language Models (LLMs) and anonymous patient data from multiple sources. This will increase the scope, speed, and quality of responses to patients' queries in their preferred language, and support clinicians to provide accurate diagnoses and treatments. They will work with end users during the design and pilot phases and verify the accuracy of the AI with human medical professionals.

João Paulo Souza of the Fundação de Apoio ao Ensino, Pesquisa e Assistência in Brazil will determine whether Large Language Models (LLMs) can be utilized as accurate information sources to guide healthcare provider decision-making. Frontline health workers must make real-life care decisions by distinguishing between relevant and irrelevant information and contextualizing it to their setting. This is particularly challenging in remote areas with limited healthcare specialists. To support them, an information program, the Formative Second Opinion (FSO), was developed to produce curated evidence summaries based on a large repertoire of real-life clinical queries. An updated version of this program is now being developed to combine a mobile messaging platform with LLMs for regions with limited internet connectivity and computer access. Using a mixed-methods study approach, they will evaluate the accuracy of the evidence summaries generated by ChatGPT-4 and those created by humans to 450 randomly selected clinical questions.

Livia Oliveira-Ciabati of the Sociedade Beneficente Israelita Brasileira Albert Einstein Hospital in Brazil will leverage AI to produce guidance and monitoring tools for less-experienced or overworked health professionals providing prenatal care via telemedicine to people from minority groups and people with greater social vulnerabilities. During the COVID-19 pandemic, Brazil's maternal mortality rate jumped to 110 deaths per 100,000 live births, which is far from their target of reaching 30 deaths per 100,000 live births by 2030. To increase the quality of prenatal care, they will integrate LLMs with an API for converting and preparing voice data and develop a simple interface to present suggestions and results to the healthcare professional. The training database will be developed using protocols defined and validated by the scientific community, including from the Global South. They will test their model with health professionals by comparing the professional decisions with the AI's suggestions which they will then follow up with a randomized clinical trial.

Alain Ndayishimiye of the Center for AI Policy and Innovation Ltd. in Rwanda will integrate a translation model with GPT-4 to produce a health service support tool in the national language, bypassing the need to build language-specific LLMs from scratch. LLMs have broad and powerful applications for improving public services such as education and healthcare by bridging information gaps across different cohorts. However, to create an impact in Rwanda, LLMs must be able to converse in Kinyarwanda, and current approaches to train LLMs in relatively minor languages are too expensive for low-resource nations. They will develop a support tool for community health workers focused on malnutrition leveraging GPT-4 as a knowledge base and an Mbaza English-Kinyarwanda translation model. The integrated support tool will be evaluated by assessing the quality of its translations.

Tamlyn Roman of Quantium Health in South Africa will use generative AI and Large Language Models (LLMs) to develop an automated analyst that integrates disparate health datasets and automates data analytics to support evidence-based decision-making in public health. Although there is a relative abundance of health-related data in South Africa, it is difficult to use effectively because the datasets are not standardized and analytics capacity to support policy- and decision-making is limited. They will source datasets for the analyst and assess the LLM's ability to automate checks and link multiple datasets. Improving interoperability between datasets will enable unique correlations to be identified between separate social indicators, which are currently recorded in distinct databases. They will also develop a user-friendly platform for output generation and visualization.

Sophie Pascoe of Wits Health Consortium (Pty) Limited in South Africa, with support from the organizations, AUDERE in the U.S. and the Centre for HIV and AIDS Prevention Studies (CHAPS) in South Africa, will develop a Large Language Model (LLM)-based application, Your Choice, that interacts with individuals in a human-like way to respectfully obtain their sexual history and improve the accuracy of HIV risk assessments to control the epidemic in South Africa. Gathering accurate sexual history is essential for assessing HIV risk and prescribing preventative drugs but is challenging due to concerns about stigma and discrimination. Your Choice, which stands for Your Own Unique Risk Calculation for HIV-related Outcomes and Infections using a Chat Engine, leverages an LLM to ensure privacy and confidentiality, improve the accuracy of risk assessments, and increase awareness of preventative treatments. This solution would provide 24/7 access to an unbiased and non-judgmental counselor for marginalized and vulnerable populations specifically, greatly reducing the barriers and concerns around seeking advice. They will co-design the app with at-risk populations and evaluate a prototype using 550 public sector healthcare providers and clients.

Theofrida Maginga of the Sokoine University of Agriculture in Tanzania will develop a ChatGPT-powered Swahili chatbot for smallholder farmers with limited literacy and scarce resources in Tanzania to detect crop diseases quickly and easily. Maize is one of the most important crops in Tanzania and generates up to 50% of rural cash income. Several diseases that afflict maize are hard to detect visually, leading to substantial losses in crop productivity and income. They will integrate AI with Internet of Things (IoT) technologies that use non-invasive sensors to monitor the non-visual early indicators of diseases, including volatile organic compounds, ultrasound movements, and soil nutrient uptake. They will also develop and integrate a Swahili chatbot to interact with farmers in their local language in a culturally-sensitive manner and perform model validation and field testing.

Shashi Jain of the Indian Institute of Science in India in collaboration with Uma Urs from Oxford Brookes University in the United Kingdom along with colleagues from Akaike and Kotak Mahindra Bank also in India, will build a GPT-enabled AI bot called SATHI, which stands for Scheme, Access, Training, Help, and Inclusion, to deliver information on the latest government financial schemes that support sectors, like micro-enterprises and farms, to potential customers and providers in rural and suburban India. Together with several partners, they will capture data and provide context to SATHI to enable it to answer queries related to financial schemes. They will also use a translation module so it can understand voice queries and respond with an audio answer in the local language. They will perform a field test at a bank branch to compare the use of SATHI alone with a human financial expert and with semi-experts supported by SATHI. They will collect data on customer satisfaction and their follow-up actions using standard field research methodology, including oral interviews and survey questionnaires.

Henrique Araujo Lima of the Universidade Federal de Minas Gerais in Brazil will develop a tool to systematically assess the accuracy and clarity of responses generated by Large Language Models (LLMs) to common questions on maternal health to increase their value in settings with limited healthcare access. To improve LLMs, it is essential to ensure the information they provide is both reliable and understandable, and for purposes such as health, LLMs will only be successful if both healthcare providers and users are confident about their benefits. They will collect the most common types of questions about maternal health in English, Portuguese, and Urdu, and submit them to the LLM. The quality of the answers will then be evaluated by medical experts from the U.S., Brazil and Pakistan, and the readability of the answers will be evaluated by individuals and a software model.

Nirmal Ravi of EHA Clinics Ltd. in Nigeria will develop and test scalable and cost-effective ways to use large language models (LLMs) such as ChatGPT-4 to provide “second opinions” for community health workers (CHEWs) in low- and middle-income countries (LMICs). These second opinions would mirror what a reviewing physician might advise the provider in question after seeing or hearing their initial report. If LLMs can enhance the capabilities of CHEWs in this way, it could improve patient outcomes, free high-skill providers for other tasks, and mitigate the serious shortage of qualified health personnel in many LMICs. The specific outcomes of this project will be: a proof of concept that LLMs can be integrated within LMIC healthcare systems to improve quality of care; a proof of concept of a system architecture for LLMs that can be scaled up and deployed progressively in LMIC healthcare systems; and an initial understanding of the capacity of current LLMs to interact with CHEWs in LMIC settings.

Imad Elhajj of the Humanitarian Engineering Initiative of the American University of Beirut in Lebanon will use Large Language Models (LLMs) to develop an interactive community health promotion platform with a chatbot that provides accurate health messages and real-time responses to queries on platforms like WhatsApp to vulnerable populations in Lebanon and Jordan. They will process texts from trusted websites, documents, and other text repositories, such as UNICEF and the WHO, into smaller text segments. These segments will then be converted into fixed-length vectors that capture their semantic meaning and contextual relationships. To generate answers, the GPT-3.5/4 model will retrieve the relevant vectors based on the user's query and use them together with the context taken from the conversation history. They will first evaluate the platform internally to ensure the relevancy, coherence and accuracy of the generated messages, and then conduct a pilot study with a small representative group from the target communities.

Maryam Mustafa of the Lahore University of Management Sciences in Pakistan will build a voice-enabled, mobile phone-based, conversational AI assistant, Awaaz-e-Sehat, for maternal healthcare workers in Pakistan to create and manage detailed electronic medical records. Pakistan has among the poorest pregnancy outcomes worldwide. The lack of documented medical records of pregnant women seeking care makes it challenging for doctors to provide accurate diagnoses and contextualized care based on socio-economic and lifestyle factors, which also play a vital role in maternal health outcomes. They will develop a proof-of-concept system comprising an intuitive user interface speech recognition module and a text recognition module to record audio responses in different languages following specific prompts. The system will then convert responses into text and populate a template electronic medical record in Urdu. Awaaz-e-Sehat will be evaluated by maternal healthcare workers at Shalamar Hospital for its ability to collect records from 500 patients.

Praveen Devarsetty of the George Institute for Global Health in India will integrate an LLM into their SMARThealth Pregnancy application to enable two-way communication support for frontline health workers to improve healthcare services for pregnant and postpartum women in India. Reducing maternal and newborn mortality and morbidity is a global priority, particularly in low- and middle-income countries where information about medical conditions and pregnancy symptoms is difficult to access in simple terms and local languages. Together with experts, they will create an "encyclopedia" of pregnancy advice based on Indian and WHO guidelines, integrate ChatGPT-4 into their SMARThealth Pregnancy application, and evaluate the application for providing high-quality and contextually relevant healthcare information and services following prompts from healthcare workers.

Anne Njeri Kamere and Mamothe Makgabo of The Aurum Institute in South Africa will develop an AI-based tool for clinicians managing complex HIV cases to support them in accurately and quickly making clinical decisions, thereby improving patient health outcomes. The tool will support clinicians in determining multiple elements: the appropriate treatment regimen and management approach for opportunistic infections, the need for preventive measures, potential referrals to higher-level facilities, and the initiation of the correct antiretroviral therapy regimen. They will train a Large Language Model and integrate ChatGPT-4 so the tool can synthesize information on a case, based on manually-entered patient details, and provide interactive guidance to clinicians based on current clinical guidelines, medication lists, and research. They will engage healthcare providers to iteratively evaluate and refine the tool in different healthcare scenarios, including integration with TB prevention and treatment.

Lara Fairall of The Health Foundation of South Africa in South Africa will enhance the AI-Supported Healthcare Assistant (AISHA) prototype, which summarizes clinical recommendations from diverse evidence sources, as a public health tool supporting clinicians, policymakers, and health educators. They will incorporate a Large Language Model to track and analyze updates in healthcare guidelines, including identifying inconsistencies between them at the global, regional, and local levels. They will also expand the database of guidelines to comprehensively cover prevalent health conditions, starting with those managed in a primary care setting in South Africa. For ongoing improvement of the AISHA tool, they will create programs bringing together local stakeholders with AI and healthcare experts from the global community. Better synthesis of healthcare evidence through AISHA will improve and expedite public health policy decision-making.