The Problem: Diagnostic Gaps in Resource-Limited Settings
Malaria remains a critical health challenge in Sub-Saharan Africa. The gold standard for diagnosis, microscopic examination of blood smears, requires trained technicians and well-equipped laboratories. These resources are scarce in regions where malaria hits hardest.
In rural Uganda, healthcare facilities face several barriers:
- Shortage of skilled technicians
- Delays in diagnosis
- Subjective interpretations prone to human error
- Limited accessibility to diagnostic services
- Time constraints due to overburdened facilities
Dr. Nakasi explained the scale of the challenge: “By the time we are beginning this project sometime back in 2015, I would say that we never had any publicly available dataset that we could utilize that had been collected using a smartphone to be able to generate insights on malaria diagnosis.”
The Innovation: A 3D-Printed Smartphone Adapter
The Makerere AI Health Lab developed a 3D-printed smartphone adapter that connects to standard microscope eyepieces. This innovation transforms existing microscopes into digital imaging devices, addressing the lack of expensive dedicated imaging equipment in resource-limited settings.
The team’s approach centered on a key question: “Can we be able to leverage especially the existing diagnostics infrastructure in Africa and how can that be supported by the emerging technologies such as artificial intelligence?”
Why Build Local Datasets?
The team chose to build their own dataset through local partnerships rather than use externally available data. This decision reflected their commitment to contextual relevance.
Dr. Nakasi noted: “Context wise, as we look at the mechanism of diagnosis which is by the use of the microscope but also wanting to leverage artificial intelligence as a technology on the other end, and how the main question to us was: how can we be able to leverage especially the existing diagnostics infrastructure in Africa?”
The goal was to create a solution that solves a significant health challenge without imposing high implementation costs.
Stakeholder Engagement: Learning from Healthcare Workers
Early conversations with Mulago, Uganda’s main referral hospital, shaped the project’s direction. Healthcare workers provided feedback that went beyond technical accuracy concerns.
They wanted a tool that would:
- Reduce fatigue from manual examination
- Support parasite identification with ease
- Integrate seamlessly into existing work routines
- Be fast but not difficult to use
- Not require ICT expertise
*Dr. Nakasi emphasized this aspect: “We wanted to bring in a tool that they can easily use and be able to integrate within their work routines that is faster but again not hard to use.”
These conversations revealed additional needs beyond parasite detection. Healthcare workers highlighted the burden of manually counting pathogens and white blood cells to determine parasitemia. This feedback led the team to expand their tool’s capabilities.
Balancing Performance and Practicality
The team aimed for 99% accuracy for single-class detection with a 5-second inference time, while ensuring smartphone compatibility. However, field testing revealed important lessons about the gap between lab performance and real-world deployment.
The Environment Matters
Dr. Nakasi shared a specific example: “There was a time where collecting data using a Samsung versioned kind of smartphone and this is the data that we used for the training of the models. The moment we took that model to a setting where we had collected the data with another kind of phone, and I think in this case it was a Tecno phone, results were clearly a little bit different.”
This experience taught the team to document development environments carefully and align them with deployment contexts. They learned to update models iteratively as environments changed.
Infrastructure Constraints Drive Design
Initial versions of the tool required internet connectivity to send images for analysis. During visits to remote health clinics, the team discovered this approach wouldn’t work.
“While we were having conversations in some of the remote health clinics that we are collaborating with, there was no internet connectivity to be able to even relay an image over the internet,” Dr. Nakasi explained.
This reality led to the development of offline detection tools that run locally on smartphones. The team also had to account for unstable power supply in rural areas.
Navigating Ethical and Bureaucratic Challenges
The project required approval from medical ethics boards accustomed to established standards but unfamiliar with AI integration in healthcare. This created iterative challenges.
Dr. Nakasi described the complexity: “Medicine or health has its already set standards in which they look at ethics. Then there is this integration where we now bring ICTs or artificial intelligence to get integrated into an already kind of domain that already has ethical guidelines.”
Key concerns included:
- Data leaving health facilities created human subject ethical concerns
- Medical standards required data to remain at facilities
- Lack of existing ethical frameworks for AI in healthcare
- Questions about patient privacy protection
- Need for mechanisms to anonymize and protect personal identifying information
The team addressed these concerns through stakeholder engagement. They worked with ethics review boards to demonstrate the impact and solutions their tool provided. Dr. Nakasi emphasized: “We’re trying to navigate the ethical pathways to making sure that there are ethical guidelines in place. We’re working now collaboratively with the Ministry of Health in Uganda to be able to come up with those frameworks.”
Partnership, Not Extraction
The Makerere approach prioritized collaboration over data extraction. Health facilities participated throughout the entire development process, from ideation to validation.
Dr. Nakasi explained their philosophy: “We’ve always drawn collaborations where, in the different projects that we engage in, they have been collaborating with us not just as partners where we’re getting data, but as partners. And so, it becomes having them part of the solution holistically.”
This approach helped overcome skepticism from healthcare experts who worried about AI replacing their roles. By involving them at every stage, healthcare workers became part of the solution rather than just consumers.
Augmenting, Not Replacing Healthcare Workers
The team designed the tool to support healthcare workers, not replace them. They emphasized that the application reduces burden but leaves final diagnosis confirmation to human experts.
Dr. Nakasi addressed this concern directly: “The conversation has always been these tools are just augmenting your work. They are just supporting whatever you’re doing to slash down the burden that you’ve been having so that they lift off the weight, rather than actually replacing you.”
According to WHO guidelines, experts should not view more than 30 patients for screening. Yet many facilities in Uganda see over a thousand malaria patients daily with only one technician or sometimes none. The tool can reduce examination time from one hour to five seconds per patient.
Human Oversight at Every Stage
The team maintained human oversight throughout:
- Data collection requires human auditing
- Development phase includes expert review
- Model outputs need human verification
- Final diagnosis confirmation remains with healthcare experts
Dr. Nakasi stressed: “We need to have a human oversight at the development phase. Whatever the model is producing, we need human oversight.”
Transparency and Explainability
The team prioritized making model outputs understandable to medical experts. They wanted to ensure that AI results aligned with medical standards.
“We want to have that explained very well and understood within the medical or health domain standards so that at the end of the day whatever solution is being provided by the application can be justified by the expert,” Dr. Nakasi said.
This transparency built trust and fostered collaboration between AI experts and healthcare professionals.
Iterative Design Based on User Feedback
The 3D-printable adapter went through multiple design iterations informed by expert feedback. Healthcare workers provided input on usability, asking for designs that allow easy smartphone insertion without repeated apparatus recollection.
Dr. Nakasi noted: “Most of whatever improvements we’ve been able to accomplish within this ocular tool were usually informed by what feedback we used to receive from the experts at the different health facilities.”
User-centered design proved essential to developing tools that healthcare workers felt comfortable using.
Scaling to Other Diseases
The Makerere AI Health Lab is expanding the framework to cervical cancer, tuberculosis, and integrating telehealth platforms. The motivation comes from addressing pressing health challenges beyond Uganda’s borders.
Dr. Nakasi explained their scaling approach: “For us, that has been the motivation to the scale: How can we utilize this solution to not only just look at one disease in Uganda, but to look at a solution that can provide a solution to the endemicity of these diseases in different parts of the world?”
The team structure evolved to include experts beyond malaria microscopy, incorporating pathologists and potentially ophthalmologists as they expand to new health challenges.
Lessons for AI Development in Africa
Dr. Nakasi offered guidance for teams working on AI in Africa:
Start with Community Understanding
“It has always been one of the key considerations, I think, as we gather requirements for the AI tools before we even develop them, to be able to understand the communities we are developing these tools for. What is the community? What is their need? And is this a fit-for-solution for them?”
Engage Stakeholders Iteratively
“That requires you to do iterations of stakeholder engagements to understand these requirements and develop the solutions alongside them.”
Build Evidence for Adoption
“For us in Africa, it has been a way to appreciate artificial intelligence and its impact in the healthcare system. It has always been a question of evidence: Can we be able to create evidence that we can stand tall on and say, ‘Yes, artificial intelligence is not only for the global north, but it can also be impactful in the global south?'”
Develop Enabling Governance Frameworks
“As Africa, we need to be intentional about developing these governance frameworks, these ethical frameworks, that are going to provide an enabling environment so that as developers come up with these tools, they are able to be integrable within the existing system.”
Focus on Patient Outcomes
“As Africa, we need to be intentional about developing these governance frameworks, these ethical frameworks, that are going to provide an enabling environment so that as developers come up with these tools, they are able to be integrable within the existing system.”
Conclusion
The Makerere malaria diagnosis initiative demonstrates how system requirements can bridge vision and practice in AI development. By prioritizing community engagement, addressing infrastructure realities, navigating ethical challenges collaboratively, and maintaining human oversight, the team created a tool that serves rather than exploits communities.
The project offers a model for equity-driven AI design in Africa, showing that technical excellence and contextual relevance can coexist when communities shape the development process from the beginning.
This article is based on the second webinar of the Africa AI & Equality Toolbox, a collaboration between the AI & Equality Initiative and the African Centre for Technology Studies (ACTS) in Kenya. The African Toolbox builds upon the methodology of the Global AI & Equality Human Rights Toolbox, an initiative of Women At the Table in collaboration with the United Nations Office of the High Commissioner for Human Rights (OHCHR).