Awards

Francis Kombe of EthiXPERT NPC in South Africa will develop an AI-based platform to support African research ethics committees and clinical trial decision-making. It will build on their cloud-based, online review system RHInnO Ethics. This system is currently used to manage the entire ethics review cycle, including protocol submission and review, approval, and follow-up, with the goal of shortening the review timeline, enhancing review quality, and speeding the discovery of life-saving public health interventions. They will consult with relevant stakeholders to identify elements of ethics review that could benefit from AI. They will then identify the required structured and unstructured data, use this data to train a model based on GPT-4, and integrate the model into their existing review system. They will evaluate the new platform, comparing it with and without the AI element and assessing results from current users, including decision quality and timeliness.

Paul Macharia of the University of Nairobi in Kenya will develop an LLM-based platform to give healthcare providers real-time access to comprehensive, up-to-date, adolescent HIV information for enhanced decision-making and better patient health outcomes. To guide the project, they will establish a community advisory board, including HIV-positive adolescents, healthcare providers, and community leaders. They will interview providers to identify their current sources of this information and their unmet needs. They will then create a dataset relevant for adolescent HIV care, including medical literature, clinical guidelines, and research findings; use it to train an LLM; and develop a natural language interface for healthcare providers to interact with the LLM. They will pilot test the platform in different healthcare settings, collecting data on its impact on provider knowledge and practice.

Philippe Manyacka Ma Nyemb of Gaston Berger University in Senegal will develop AI-based approaches to better monitor and manage cardiovascular diseases and understand their risk factors. They will perform a household-level study in the Saint-Louis Region of Senegal with monthly data collection, including medical examinations, behavioral surveys, and physical environment assessments. The monthly monitoring data will be analyzed by AI-based approaches, yielding cardiovascular disease risk scores for household members. They will use the study data and risk scores to train a Large Language Model with a chatbot interface available to healthcare professionals and the public. The chatbot and data collection process will serve as an integrated platform to reduce the burden of cardiovascular diseases. It will increase awareness of the disease and its risk factors for the public, and it will help increase adoption and effective use of digital tools more broadly to improve health.

Jules Brice Tchatchueng Mbougua of Centre Pasteur du Cameroun in Cameroon will develop a chatbot to provide health information in the Fulfulde language, which is commonly spoken in West Africa, to increase equitable access to healthcare. To overcome language barriers as well as variable levels of literacy, the chatbot will interact with users by speech or text and with bidirectional translation between Fulfulde, French, and English. To enable this, they will compile an extensive Fulfulde dataset covering health-related expressions and terminology. They will develop health information content that is culturally relevant by co-creating it with the communities the tool is meant to serve, and it will be focused on primary healthcare. The chatbot will include a way for users to provide feedback to ensure it is delivering information most relevant to the evolving needs of Fulfulde-speaking communities.

Mandicou Ba of Université Cheikh Anta Diop in Senegal will develop an AI-based tool for automatic, cost-effective, and accessible early diagnosis of the eye disorder strabismus and for guiding surgical correction. Strabismus is eye misalignment, the two eyes pointing in different directions, and the associated impaired vision can become permanent at a young age if uncorrected. They will collect a clinical dataset of facial images of strabismus patients in Senegal, annotated by experts. After identifying a suitable AI-based method, they will apply it to the dataset to create an AI model for diagnosis and accurate calculation of angular deviation between the two eyes for use during surgical repair. They will use the model to develop an automated system as a web-based tool and also as a smartphone app, making it accessible even in rural areas for early diagnosis in children.

Mamadou Bousso of Iba Der Thiam University in Senegal will develop methods for AI-based analysis of ultrasound images for cost-effective early detection of liver fibrosis caused by hepatitis B viral infection. They will improve the performance of an existing method by acquiring ultrasound data that more comprehensively encompasses the clinically-recognized stages of liver fibrosis. The expanded dataset will be used to train an AI model well-suited to capture complex patterns in imaging data. They will also establish support for healthcare professionals that facilitates the adoption and effective use of the application, including training courses, web-based and mobile phone-based tools with user-friendly interfaces, and ongoing technical support. The application would enable more screening in underserved areas, with increased early detection and awareness of liver fibrosis decreasing mortality from the disease as well as healthcare costs.

Mame Marème Fall of Kajou Senegal in Senegal will develop a platform to increase access to accurate health information, including information on available healthcare services, to improve the quality of life for rural populations in Senegal. They will use a Large Language Model to create a database of health information with a chatbot interface enabling questions and answers by either speech or text, including spoken questions in either French or Wolof. The database will be accessible online via internet technology that accommodates connections of short duration and low bandwidth, and it will be available offline as content stored on mobile phone microSD cards. They will evaluate the quality of answers to health questions through the system by engaging experts and by surveying users through pilot distribution of 1,000 microSD cards with the database.

Yap Boum II of Institute Pasteur of Bangui will develop a digital platform, called The Village, that strengthens the scientific research capacity across the Pasteur Network and beyond through conversational chatbots that forge productive links between those seeking and offering resources, ideas, and collaboration. They will identify suitable Large language Models and create a chatbot that collects unstructured data through conversations with scientists to generate user profiles with higher potential for productive matching across the research community. They will test and continually refine the platform through in-person and virtual meetings across the scientific community. As a platform making connections between scientists regardless of their location, resources, and research capacity, The Village will increase equity in global health research.

David Kamau of Mary Help of the Sick Mission Hospital in Kenya will integrate ChatGPT and a medical Large Language Model with the existing health management information systems in Kiambu County, Kenya, to provide clinical decision support for noncommunicable diseases. This integration will support healthcare providers in diagnosing diseases accurately and quickly, reducing misdiagnoses and improving patient outcomes, and it will support optimizing treatment plans, reducing unnecessary procedures and healthcare costs. They will evaluate the performance of the integrated AI tools, assess the usability of the system through surveys, and capture demographic data for patients receiving AI-assisted care. They will also provide training for healthcare professionals on effectively using AI tools to enhance patient care.

Steven Wanyee of IntelliSOFT Consulting Limited in Kenya will develop an AI-based framework for analysis of cervical cancer registry data to identify epidemiological trends and improve surveillance and prevention efforts. The analysis will incorporate variables such as demographic factors, geographic locations, screening history, HPV vaccination rates, and treatment outcomes. They will use natural language processing to extract and analyze unstructured data. Machine learning algorithms will be used to identify patterns and trends in cervical cancer incidence rates, stage at diagnosis, treatment outcomes, and survival rates. They will develop predictive models to forecast cervical cancer burdens, estimate the potential impact of interventions for prevention, and guide resource allocation and targeted prevention strategies. They will also create user-friendly interfaces and visualizations to enable policymakers, public health professionals, and researchers to easily interpret the modeled data and use it effectively.

David Kamugundu of eFiche Limited in Rwanda will develop an AI-based platform to support nurses in Rwanda in accurate and efficient diagnosis and patient treatment. The platform will be integrated into the already-operational, web-based electronic medical records system eFiche. It will serve as a virtual assistant for nurses, including providing diagnostic support, suggesting treatment plans based on the latest medical guidelines and assessing adherence, and identifying potential adverse events. They will train a Large Language Model with historical health data to help predict diagnostic outcomes, laboratory orders, and subsequent procedures and prescriptions, and they will develop a chatbot to provide information and recommendations to nurses in a conversational manner. They will evaluate the platform based on the accuracy of the model, satisfaction of users, and the impact on healthcare delivery, including diagnostic accuracy, treatment efficacy, and adherence to health guidelines.

Obed Brew of Kwame Nkrumah University of Science and Technology in Ghana will explore applying a combination of AI-based analytical approaches to clinical data for early detection of preeclampsia in pregnancy to reduce maternal and neonatal morbidity and mortality. They will collect a diverse set of data from pregnant women, including physiological data from wearable devices, electronic health records, clinical notes, and fetal nucleic acids from non-invasive prenatal testing. They will apply a variety of AI tools to detect patterns across the data, such as associations between fetal gene signatures and maternal physiological markers. While evaluating the performance of the AI models in detecting preeclampsia, they will develop training programs for healthcare professionals in the use of AI tools in clinical settings.

Mihaja Raberahona of Equipe de Recherche Clinique en Maladies Infectieuses in Madagascar will develop an AI-based triage tool to identify patients likely to have pulmonary TB based on combining physiological data from wearable devices, cough acoustics, and anthropometric data. They will create an AI model that can find patterns in the combined data to help staff with minimal medical training quickly prioritize patients needing confirmatory testing. From a cohort of patients with symptoms suggestive of TB from primary healthcare centers in Madagascar, they will collect physiological and anthropometric data and perform a digital chest X-ray, cough sound analysis, and a GeneXpert molecular assay for the presence of Mycobacterium tuberculosis. They will evaluate performance of the AI-based triage tool primarily as compared to the microbiological assay, but also to other current screening methods, including chest X-rays analyzed by clinicians with or without AI support.

Jesse Gitaka of Mount Kenya University in Kenya will develop the GPT-4 AI model to support healthcare providers with up-to-date medical information for improved clinical decision-making and neonatal care. They will identify knowledge gaps among providers, using this information to guide training for a cohort on using GPT-4 for clinical support. They will then evaluate their use of GPT-4 for its impact on clinical decisions and neonatal outcomes. A subset of the healthcare provider cohort will be engaged to identify the barriers, risks, and opportunities associated with use of the AI tool. Results from both these evaluations will be used to develop a scalable framework for the deployment of AI in similar healthcare contexts to reduce neonatal morbidity and mortality, especially where there is a shortage of medical personnel.

Fred Kaggwa of Mbarara University of Science and Technology in Uganda will develop an LLM for answering questions related to prenatal and postpartum care in Uganda in three languages: Swahili, Runyankore-Rukiga, and Luganda. This will also serve to create a scalable, open-source pipeline for developing AI models incorporating underrepresented languages more broadly. They will identify the most suitable LLM and train it with existing databases and with new, high-quality medical text, including that extracted from textbooks and patient education materials as well as new question-and-answer pairs written by local clinicians. They will iteratively evaluate and improve the LLM, working directly with community health workers and expectant mothers to ensure the model’s responses are safe, accurate, relevant, and accessible.

Dennis Maorwe of DPE Company Limited in Kenya will develop a Large Language Model (LLM) to support the planning and execution of behavior change interventions to improve health outcomes for Kenyans living with SCD. Guided by insights from SCD patients and medical specialists, they will train an LLM with SCD management practices and additional open data relevant to designing interventions for social behavior change for SCD. The LLM will be used to generate tailored health messages to support SCD patients through the different disease management stages and to empower them to take an active role in their health. This will help reduce the stigma around the disease and improve their quality of life.

Brian Turyabagye of MamaOpe Medicals in Uganda will develop an AI-based platform integrated with the MamaOpe screening tool to enhance the diagnosis and management of respiratory illnesses. MamaOpe is a pneumonia screening tool based on vital signs, including respiratory rate and lung sounds, with results presented via a mobile app into which healthcare workers can enter patient symptoms and store patient histories. They will train a Large Language Model with the MamaOpe database to help guide differential diagnosis by highlighting critical indicators and suggesting supplementary tests for more accurate diagnosis. In settings with limited direct access to specialists, the platform can facilitate remote consultations. They will perform an efficacy study to evaluate how well the platform correctly predicts clinical outcomes.

Rose Nakasi of Makerere University in Uganda will develop an AI-based platform to support diagnosis and management of cervical cancer in Uganda. They will collaborate with the Uganda Cancer Institute to develop a set of AI tools for automated diagnosis of cervical cancer based on microscopy of patient samples and for automated generation of the associated pathology reports. This will include categorizing pathologies, enabling identification of trends over time. The AI tools will be integrated into a web-based platform along with the capacity for Visual Question Answering to support interpretation based on medical images and diagnosis in remote areas of the country with limited access to pathologists. They will evaluate the accuracy of the cervical cancer diagnoses and the quality of reports generated through the platform as compared to those generated by expert pathologists.

Kevin Korir of Visortech Solutions in Kenya in partnership with Yemaya Health Advisory will develop an LLM to map the process of health policy creation and approval, serving as a tool for quicker translation of new evidence into policies. They will pilot test the tool by using it to develop a policy framework for strengthening national HIV prevention management systems. They will develop the LLM as support for policy simulation and scenario planning to inform the policy approval process, for gap analysis to optimize current policies, and for identification of conflicts across related policies and their implementation. The LLM will facilitate generating written documents useful for policymakers, including briefs for new policies and summaries of proposed changes to existing ones.

Prince Adjei of Kwame Nkrumah University of Science and Technology in Ghana will develop an AI-based tool to support clinicians and patients in managing complications and comorbidities in pregnancies, focusing on pre-gestational and gestational diabetes. They will train a Large Language Model with relevant data and health guidelines from Ghana, including a glossary of medical and nutritional terms in the local language Twi. They will also design a user-friendly interface for interactions in English and Twi, with the patient interface incorporated into WhatsApp. The resulting tool, RAID-MaP-Gh, will provide practical guidance to clinicians, ranging from specialists to midwives and traditional birth attendants, and guidance to patients that takes into consideration the socioeconomic and cultural contexts most relevant to Ghana. The tool will help close health information gaps associated with gestational diabetes, reducing the number of undiagnosed cases and improving maternal and infant health outcomes.

Polly Okello of Medbook Kenya Limited in Kenya will develop a set of LLMs to support frontline healthcare workers in rural areas and among marginalized populations in Kenya. The LLMs will be based on the clinical LLM Med42 and trained with information representative of the Kenyan healthcare system, including clinical data and guidelines. The LLMs will be used to provide information to clinicians about medical conditions, treatments, and medications; summaries of patient records and other medical documents; personalized treatment plans for patients; and educational materials for patients and healthcare workers. They will deploy the LLMs and evaluate their impact through surveys with healthcare workers and patients and through case studies. They will assess the knowledge and skills of frontline healthcare workers, quality of patient care, efficiency of healthcare delivery, and patient satisfaction, with the goal of making healthcare more efficient, effective, and accessible in Kenya.

Andrew Maranhão Ventura Dadario of Hospital Israelita Albert Einstein in Brazil will test and evaluate LLMs for their ability to structure and anonymize clinical notes. Such structured notes would promote data interoperability, which would improve patient care, research, and health policies for the diversity of Brazilian public health patients; and establishing standardized evaluation of LLMs for public health would promote their future applications. They will prepare a representative dataset by extracting, processing, and annotating clinical notes, and then test and compare a set of existing LLMs in extracting information from text, including their ability to preserve patient privacy, their applicability to the Portuguese language, and their suitability to primary healthcare in Brazil encompassing equitable performance across patients stratified by gender, race, ethnicity, and age. The best performing LLM will be used to transform free text from clinical notes, and the structured product will be assessed by health managers for accuracy and usefulness.

Madji Sock of Haskè Ventures in Senegal will develop an AI-based platform for patient intake in Senegalese hospitals that will streamline intake and monitor disease trends to improve healthcare delivery and outcomes. The platform will use a medical Large Language Model (LLM) to help hospital reception staff make initial assessments of patients for quick and accurate triage. The LLM will be adapted for local dialects, medical practices, and patient demographics, and it will interact with patients by voice and text display. Daily automated analysis of patient data from multiple hospitals will identify disease patterns and trends, enabling quicker response to disease outbreaks. They will pilot test the platform in selected hospitals and health centers, including training for staff and continuous evaluation and improvement of the platform.

Hugo Fernandes of Universidade Federal de São Paulo in Brazil will create a university center focused on using AI to support primary healthcare professionals in surveillance and clinical decision making for complex cases of interpersonal violence in vulnerable populations in Brazil. AI systems could help in diagnosis and classification, risk determination, recommendation of interventions, referrals to relevant care providers, socio-emotional support plans, support for perpetrators in avoiding repeated offenses, mandatory notifications, as well as generalization across contexts for improved understanding of patterns of violence. They will aggregate information technology at a campus site, train a Large Language Model (LLM) with appropriate data and create a prototype tool, and gather relevant experts to judge the quality of the tool for its potential contributions to Brazil’s health system in targeting interpersonal violence.

Cecília Claudia Costa Ribeiro of Universidade Federal do Maranhão in Brazil will develop an LLM to help frontline healthcare workers identify risks and take action to prevent NCDs in the first 1000 days of life. There are connections between NCDs in early childhood that are affected by social inequities as well as pregnancy-associated factors. They will apply machine learning algorithms to data for women during and shortly after pregnancy to build risk prediction models of NCDs for children in the BRISA Cohort of São Luís, with predictor variables encompassing socioeconomic, behavioral, and biological stressors. The estimated risks will be input to an LLM that incorporates the best available scientific evidence on each NCD. This LLM, when given data from women patients, can then determine the risk of NCDs for the patient’s children, along with an explanation to help CHWs effectively communicate these risks and make evidence-based recommendations in real-time for prevention of NCDs.

Elisdete Maria Santos de Jesus of Universidade Estadual de Campinas in Brazil will train an LLM to read Brazilian medicine package leaflets and generate accurate, up-to-date summaries of how to use the medicine that is easier to understand for patients who take the medicine and for healthcare professionals who prescribe, dispense, and administer it. They will do a literature review, including articles in English and Portuguese, to better understand the types of information in leaflets and how understandable it is to users of the medicine. Subsequently, medicine leaflet information will be collected, processed, and used to train an LLM to create summaries in plain language, after which the summaries will be evaluated for their quality. The trained LLM could be used to help develop additional health education tools.

Momy Seck of Station d'Innovation Aquacole in Senegal will apply AI approaches to remote sensing data to map spatiotemporal changes in the risk of the neglected tropical disease schistosomiasis and automatically generate public health bulletins supporting geographically-targeted control of the disease. Schistosomiasis, also called bilharzia, is caused by a parasitic worm transmitted by aquatic snails. The snails’ presence can be monitored by remote sensing by virtue of their association with a particular aquatic plant. They will use machine learning to better analyze the remote sensing data and generate schistosomiasis risk maps. They will then use Vision Language Models to extract information from these maps to be used by Large Language Models to automatically generate bulletins on risk. They will refine the bulletins through workshops with relevant public health authorities to assess their needs and the usefulness of the bulletins.

Zilma Silveira Nogueira Reis of Universidade Federal de Minas Gerais in Brazil will develop a Large Language Model (LLM) that can generate personalized instructions for taking medications that are tailored to a user’s level of literacy, cultural context, and special needs to increase equitable access to medications and promote safe and effective patient self-care. They will determine the contexts that medicines are prescribed in the Portuguese-speaking countries Brazil, Mozambique, and Portugal; and they will train an LLM with a Brazilian public corpus of medical leaflets, social media text, and other relevant documents. A team of prescribing professionals will evaluate the LLM product for use as a web-based platform, including its ability to use an individual patient profile to generate personalized text instructions with supporting pictograms, such as those indicating relevant body parts, dose, unit of measurement, and mechanism of administration.

João Paulo Papa of Sao Paolo State University in Brazil will develop a platform in which Large Language Models (LLMs) will extract information from published research articles and generate one-page fact sheets highlighting the potential implications for public health policy in an accessible form for researchers and health system managers. They will integrate into the LLM training a methodology for knowledge translation that has been used to train health system researchers to communicate results in a form they can most readily be used.

Sandra Maria do Valle Leone de Oliveira of Fiocruz in Brazil will develop an early warning system for infectious diseases that integrates community leader engagement and a Large Language Model (LLM)-based tool for health-related event-based surveillance in two Brazilian cities on the Bioceanic Route transcontinental railway. Monitoring information from formal and informal internet sources about health-related events can lead to quicker detection and response for disease outbreaks. They will develop the system with leaders across diverse roles in the community, who will advise on the list of reportable events and priority health conditions and be trained in reporting through the system. They will create an online reporting form and a process for escalating alerts to local or national health authorities for analysis to quickly enable decisions about potential responses to disease outbreaks. Data accumulated through this system will be used to train an LLM to extract patterns and generate accurate and relevant predictions.

Shirley da Silva Jacinto de Oliveira Cruz of Universidade Federal de Pernambuco in Brazil will develop an LLM-based chatbot and use it in a pilot study to support post-discharge follow-up communication with surgical wound patients to promote a faster and safer recovery. A chatbot would provide a personalized and accessible mechanism for overcoming geographical, socioeconomic, and communication barriers between the healthcare team and the patient, including care reminders and regular check-ins via text and voice to acquire data on the patient's recovery and alert healthcare professionals to signs of complications. They will survey the needs at the Hospital das Clínicas de Pernambuco pilot site, train an LLM using electronic patient records from the hospital, and perform a series of evaluations using feedback from surgery patients and hospital staff to assess the usefulness and relevance of the information the chatbot provides.

Ida Viktoria Kolte of Fiocruz in Brazil will adapt ChatGPT for use in the languages of indigenous populations in Brazil (Guarani-Kaiowá communities) and Mozambique (Echuwabo-speaking communities in Nicoadala) to support community health workers (CWAs) in reducing the burden of tuberculosis. Translating information on tuberculosis from Portuguese would enable its integration with the traditional medicine practiced by indigenous populations particularly vulnerable to the disease, helping overcome language and cultural barriers and a history of marginalization. They will train a translation model using existing translated texts and validated common questions and answers about tuberculosis in the literature. They will test and improve the model through assessment of its accuracy and cultural appropriateness by CHWs, traditional medicine practitioners, and community members; and they will perform a two-month pilot trial of its use by CHWs in Brazil and Mozambique.

Renee Street of the South African Medical Research Council in South Africa and Samuel Scarpino of Northeastern University in the U.S. will perform a proof-of-concept study of AI for wastewater-based epidemiology by fine-tuning Large Language Models (LLMs) with SARS-CoV-2 metagenome sequence data to detect the emergence of viral variants. Monitoring wastewater is a useful surveillance tool that encompasses infected individuals with varying disease severity and access to healthcare facilities. They will fine-tune two LLMs to identify the emergence of viral variants using two existing SARS-CoV-2 metagenome sequence data sets: data processed with established bioinformatics to identify variants and data collected from wastewater before, during, and after the emergence of the Omicron variant. Success of the models will help validate an AI-based approach for monitoring microorganisms in wastewater to track the prevalence of specific health threats and forecast disease outbreaks, enabling targeted public health interventions.

Nicolas Poussielgue of Dakar Institute of Technology in Senegal will expand the Weerwi platform by incorporating a Large Language Model (LLM) into the existing chatbot to enable more personalized advice and information on menstrual health for young women. The LLM will enable interactions in a conversational style to better engage users. It will also incorporate conversations across engagements with individual users to better identify their priorities and gaps in knowledge and give personalized recommendations for where to find additional information. The enhanced chatbot will be designed in consultation with users, and like the existing version, it will be available through a website and mobile app.

Ziyaad Dangor of the Vaccines and Infectious Diseases Analytics Research Unit of Wits Health Consortium (Pty) Limited and Jenny Coetzee with Minja Milovanovic of African Potential Management Consultancy (Pty) Ltd, together with Andreas Cambitsis and Emma Gibson of Business Science Corporation Global and Mirjam van Reisen and Samson Yohannes of Virus Outbreak Data Network (Vodan)-Africa, all in South Africa, will develop an LLM supporting clinicians’ note taking to improve patient care. The LLM will be integrated into 1beat, a South African-developed electronic medical record system. They will fine-tune the LLM as an AI assistant during patient intake and discharge to improve quality as well as consistency in digital patient records, and they will ensure the clinical data is Findable, Accessible, Interoperable, and Reusable (FAIR). This will improve patient care as well as health research. They will also survey relevant stakeholders to develop a business model to support the further development and sustainability of the 1beat system.

Neil Martinson of Perinatal HIV Research Unit of Wits Health Consortium (Pty) Limited in South Africa will apply AI-based approaches to support increased screening for high-risk HPV and its treatment to reduce the morbidity and mortality associated with cervical cancer. The scale-up of cervical HPV screening will lead to more cervical biopsies as part of follow-up care and treatment where indicated. Automated approaches for biopsy analysis would overcome the limited number of specialist pathologists otherwise required. They will train an LLM with pathology reports of cervical biopsies to extract critical wording, and they will apply computer vision for analysis of the matching digital images from each patient’s pathology slide. They will then develop a machine learning algorithm for automated reporting that accurately separates normal images from those with pathology, categorizes the pathology, and determines if the pathology extends to the biopsy margins, which indicates it is likely to recur.

Jennifer Chipps of the University of the Western Cape and Damian Clark of the University of KwaZulu-Natal in South Africa will apply machine learning to a hospital digital registry of trauma and surgical patients to develop an algorithm for predicting unplanned and avoidable readmissions to improve patient outcomes and reduce the burden on the healthcare system. Unplanned hospital readmission within 30 days is an important indicator of the quality of patient care. They will use de-identified patient data from a South African public hospital to train an AI model to predict readmissions avoidable under the current standard of care. The resulting algorithm will be validated using real-time patient data, integrated into the hospital workflow as a Readmission Prediction Classifier Tool, and tested in different clinical settings. The tool will enable treatment plans tailored to individual patients and their risks, improving health outcomes and reducing financial burdens for patients and healthcare providers.

Kahesh Dhuness and Charita Bhikha of the Council for Scientific and Industrial Research in South Africa will use statistical analysis and machine learning techniques on existing data to increase the effectiveness, as a triage tool, of Umbiflow, a Doppler ultrasound device used for antenatal screening. Umbiflow monitors umbilical artery blood flow to help determine the risk of impaired fetal growth. Statistical analysis of Umbiflow data will identify patterns of fetal risk across clinics and hospitals to help guide effective resource allocation. Application of machine learning to Umbiflow data from clinical trials will enable identification of complex patterns to improve determination of fetal risk and subsequent triage. They will also integrate a Large Language Model to support healthcare providers in evidence-based decision making using Umbiflow data.

Rahel Mekonnen with Hiwot Mekonnen of the Ethiopian Artificial Intelligence Institute in Ethiopia will develop a multilingual speech-to-speech translation system to remove language barriers in communication between healthcare providers and their patients. This system would use three of the local languages: Amharic, Afan Oromo, and Somali, making healthcare less dependent on the availability of translators, which is a challenge due to the multiple languages spoken in Ethiopia, and enabling direct and open communication with patients without the need for an intermediary. They will customize an existing LLM by training it with medical text data and transcribed audio data from anonymized doctor-patient conversations, including text-to-speech model training using text read by language experts. They will create both a mobile phone-based translation application for patients and a web-based application for physicians.

Martha Yifiru with Dinksira Bekele of Addis Ababa University in Ethiopia will develop a smartphone-based conversational agent in the Amharic language to provide information and answer questions for women on pregnancy, childbirth, and their children’s health. This could reduce maternal and neonatal mortality by providing advice on best practices, including sensitive discussions in private on personal health; detecting early warning signs to enable prevention of health complications for mother and child; and ensuring timely visits to healthcare facilities while avoiding unnecessary visits. They will identify a suitable medical LLM, adapt it for use in Amharic with translations of relevant manuals and other documents, develop the agent with its interfaces between the LLM and user, and validate that the agent’s healthcare advice is as expected, with the possibility of expanding the number of users by including additional Ethiopian languages and creating a speech interface.

Lynn Hendricks with Simone Titus and Karolien Perold-Bull of Stellenbosch University and Annibale Cois of the South African Medical Research Council, all in South Africa, will use multiple AI applications to develop a culturally sensitive, visual storytelling intervention for patients that supports their communication with healthcare providers and enhances their knowledge and management of their health status and care regimen. With patients and healthcare providers, they will collect high-quality evidence-based data, representative of the African context and related to symptoms, diagnosis, patient experiences, treatment plans, and disease information. This data will be used with a set of AI tools to generate a repository of text stories with associated images and short videos. They will then recruit a cohort of patients with a new diagnosis of a non-communicable or infectious disease and test whether giving them personalized relevant stories in various formats increases their understanding of their diagnosis and treatment plan.

Mbaye Faye of Tech Care For All – Africa in Senegal will develop Large Language Models to support continuing education for healthcare professionals and their mastering of critical guidelines and procedures for patient care. The Medical Learning Hub and its mobile app already supports hospitals and healthcare training organizations as a tool to track training attendance, host content, distribute certificates, and conduct assessments. They will integrate ChatGPT to transform static patient-care guidelines into dynamic, interactive learning modules, enabling real-time updates as guidelines evolve and more effective e-learning. They will pilot the approach through the Continuous Medical Education processes of four large hospitals in Senegal. Their platform will serve healthcare organizations in overseeing staff competencies and compliance with standards, and it will foster the professional development of healthcare practitioners for enhanced patient care.

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.

Scott Mahoney and 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.

Mariam Parker of the University of the Western Cape in South Africa will develop an AI-based application, MzansiMed, for real-time translation of South African languages to increase healthcare access and improve health outcomes. To train an AI model, they will build a database of dialogues encompassing a variety of healthcare scenarios, initially collecting data in at least two widely spoken languages, including isiXhosa and Zulu. This will be in collaboration with linguists and participant communities to ensure incorporation of linguistic diversity, cultural nuances, and colloquial expressions. They will take a user-centric design approach to create an intuitive user interface for interactions by speech or text, integrating ChatGPT-4 and including visual aids. They will iteratively refine the app, including testing it with patients and healthcare providers, to ensure the translation is accurate, effective, and culturally sensitive.

Candice Chetty-Makkan and Sophie Pascoe of the Health Economics and Epidemiology Research Office of Wits Health Consortium (Pty) Limited in South Africa will develop an AI-based mobile phone application for use after HIV self-testing to increase access to counseling and advice on seeking healthcare services. They will build on an existing Large Language Model-based chat interface named Your Choice, which stands for “Your Own Unique Risk Calculation for HIV-related Outcomes and Infections using a Chat Engine.” The new app will have improved conversational abilities and generate conversation summaries for patients and healthcare providers. They will pilot test the app using a hypothetical HIV test result, engaging a subset of individuals and healthcare providers previously recruited for their insights on health-seeking behavior. The app will be integrated into the HealthPulse AI mobile app platform, providing a more accessible alternative to facility-based HIV testing and the associated guidance on next steps.