Makerere AI Health Lab: Using smartphone-based AI to diagnose malaria in Uganda's health centres

The problem

Uganda has the highest malaria incidence in the world — 478 cases per 1,000 people per year. It is the country's leading cause of sickness and death. Since January 2022, there has been a renewed wave of the disease, with a peak of more than 300,000 casesreported every week.

Diagnosing malaria accurately depends on microscopy — examining a blood sample under a microscope to identify the parasite. This is considered the gold standard because it provides a definitive diagnosis and can identify the specific species causing the infection, thereby determining the appropriate treatment. The alternative, rapid diagnostic tests, are quicker but miss infections when parasite levels in the blood are low.

The problem is that microscopy requires trained laboratory technicians. In most health centres, particularly in rural areas, the few technicians available are responsible for processing a volume of samples that far exceeds what is sustainable. The World Health Organisation recommends that a microscopist examine no more than 30-40 slides per day to maintain quality. In practice, this limit is regularly exceeded because the demand is too high.

Each slide must be examined by hand — fixed, stained, and then studied under the microscope around 100 times before a diagnosis can be made. An experienced technician takes an average of 15 minutes per slide. The physical strain of hours spent looking through a microscope leads to fatigue, and fatigue leads to errors. A wrong malaria diagnosis can mean a patient receives the wrong medication, which, in the worst case, results in drug resistance or death.

The result is a system where the most accurate diagnostic method is too slow and too dependent on scarce human expertise to keep pace with the scale of the disease. Patients queue for results. Technicians are overstretched. And diagnostic quality suffers as the workload increases.

The solution

The AI Health Lab at Makerere University in Kampala found a way to make malaria diagnosis faster and more consistent without replacing the equipment or the people already doing the work.

The setup is simple. A smartphone is attached to a standard microscope — the kind already found in health centres across Uganda — using a small 3D-printed adapter designed by the lab. When a technician places a blood sample slide under the microscope, the smartphone captures a digital image. AI then analyses that image, identifying and marking individual malaria parasites.

The AI has been trained on thousands of microscopy images collected from real health centres and annotated by laboratory professionals. It recognises the visual patterns associated with malaria infection. But it does not make the diagnosis on its own — it presents its findings to the technician, who reviews and confirms the result.

The 3D-printed adapters were tested in the field and adjusted based on feedback from health workers. A companion app captures each microscopy image along with details about the sample and the microscope settings, building a growing dataset that improves the AI over time.

Because the approach is built around a smartphone and a microscope adapter, the same setup can be retrained for different diseases. The lab is now piloting it for cervical cancer screening with the Uganda Cancer Institute and for tuberculosis diagnosis at five regional referral hospitals. They are also exploring its potential for sickle cell disease and anaemia.

Beyond diagnosis, the lab is building tools to help with what happens after a result. A real-time surveillance system combines diagnostic data with environmental information — rainfall patterns, mosquito breeding conditions, bed net coverage — to help forecast outbreaks and target public health responses geographically. And AI-powered chatbots in local Ugandan languages are being developed to help communities understand malaria prevention, symptoms, and treatment.

The project is led by Dr Rose Nakasi with a team of 22 researchers, lab technicians, and pathologists, funded by Google, the US National Institutes of Health, the Lacuna Fund, and other sources.

The impact

1. A diagnosis that took 15 minutes now takes 90 seconds

An experienced technician examining a blood smear by eye spends an average of 15 minutes per slide — carefully studying the sample under the microscope, looking for the small parasites that indicate malaria. The AI-enhanced system does this in approximately 90 seconds. For a health centre seeing dozens or hundreds of patients a day, that difference transforms how quickly people receive their results and how soon they can begin treatment. It also means technicians spend less time straining their eyes through a microscope, reducing the physical fatigue that accumulates over a working day and contributes to diagnostic errors.

2. Accuracy of up to 99%

The AI matches or exceeds the accuracy of experienced human microscopists. This matters most when parasite levels in the blood are low — exactly the cases that rapid diagnostic tests often miss and that tired technicians are most likely to misread. Unlike a person, the AI does not lose accuracy as the day wears on. It applies the same level of attention to the last slide of the day as it does to the first. In a disease where a wrong diagnosis can mean the wrong medication, drug resistance, or, in the worst case, death, that consistency has direct consequences for patient safety.

3. 60% more patients were diagnosed in the same time period

Health centres using the tool can process significantly more samples without adding staff or extending working hours. By handling the most time-consuming part of the process, the AI helps health centres stay within the WHO recommendation of processing a maximum of 40 slides per day while still meeting patient demand. For a country with the world's highest malaria incidence and a severe shortage of laboratory technicians, a 60% increase in diagnosis without sacrificing accuracy is not an incremental improvement — it changes how many people receive a reliable diagnosis.

4. Tested across approximately 40 government health centres

The system has been piloted in real clinical settings — government health facilities serving real patients — not just in university laboratories. The 3D-printed adapters, the data collection app, and the AI models were all refined based on feedback from the health workers who used them in their daily routines. This means the evidence for the tool's performance comes from the conditions it would need to work in at scale: busy, potentially under-resourced health centres with high patient volumes and limited staff.

5. The same approach is being applied to cervical cancer and tuberculosis

Cervical cancer is the leading cause of cancer death among Ugandan women. Uganda has the seventh-highest cancer incidence rate worldwide. Tuberculosis kills thousands each year, with 32% of the approximately 91,000 people who fall sick with TB in Uganda annually also living with HIV. Both diseases are diagnosed through microscopy, and both face the same constraints — too few trained technicians, too many samples, too little time. The lab is now piloting the same smartphone-and-microscope approach for cervical cancer screening in partnership with the Uganda Cancer Institute and for TB diagnosis at five regional referral hospitals. If the method proves effective across these diseases, a single infrastructure investment — smartphones, adapters, trained AI models — would serve multiple diagnostic needs rather than requiring separate solutions for each.

6. Full national integration has not yet happened

The tool has been piloted, and the results are strong. But formally integrating it into Uganda's national Health Management Information Systems requires approval from the Ministry of Health, which is still in progress. Currently, when a patient is diagnosed at a health centre, the result can take between a week and a month to reach the national system through manual reporting. Integration would mean results are recorded electronically immediately after diagnosis — giving the Ministry of Health a far more timely and accurate picture of disease patterns across the country. Until that approval comes through, the tool operates within individual health centres rather than as part of the national health infrastructure.

Lessons learned

• Augmenting existing equipment lowers the barrier to adoption. The decision to build around microscopes and smartphones already present in health centres meant adoption did not depend on new procurement or infrastructure. For governments considering AI in service delivery, whether in health, inspection, or any field-based work, the question of whether a tool requires new equipment or can work with what staff already have often determines whether it is used beyond the pilot.

• How staff perceive the tool matters as much as how well it performs. Positioning the AI as something that lightens the technician's workload, rather than as something that could replace them, shaped how it was welcomed. For any government introducing AI into the professional workforce, the framing, and the design choices that back it up, such as keeping the human in the decision-making loop, affect adoption as much as the technology's accuracy.

• Planning for data collection from day one creates a compounding advantage. The lab designed data collection into the clinical workflow from the start, building a dedicated app that captures images and metadata during routine diagnostics. Every diagnosis contributes to a growing dataset that improves the models over time. Organisations that treat data collection as a separate, later step often find they lack sufficient high-quality data when they need it.

• Reaching the hardest-to-serve facilities depends on more than accuracy. The strong early results came from selected health centres, and the harder challenge lies in reaching remote areas, where running and supporting the tool is more demanding. Patients used to the traditional method also have to be persuaded to accept a smartphone-based result. For a government planning to scale a field diagnostic, the conditions in the least-connected settings and the confidence of the people relying on the tool shape how far it spreads, not accuracy alone.

This case study was written with assistance from artificial intelligence.