An AI-assisted febrile illness surveillance system piloted across four Ugandan districts has recorded high agreement with the judgments of Village Health Teams, and is now moving into a second phase that will test something considerably harder: whether showing health workers the AI’s interpretation improves their decisions or simply displaces them.
The workflow is built into Uganda’s electronic Community Health Information System and running in Mpigi, Moyo, Buikwe and Lira. It is implemented by the Ministry of Health with Medic, using Audere’s HealthPulse AI Interpreter to provide AI-assisted interpretation at the point of community-level care.
Why the first phase was designed the way it was
Throughout Phase I, the AI produced an interpretation of each case — but Village Health Teams could not see it. The system ran in parallel, silently, while frontline workers made their own assessments unaided.
That design choice is the most important detail in the study. It means the recorded agreement between AI and human interpretation is a genuine comparison rather than a measure of how readily workers deferred to a machine. Researchers got a clean baseline before introducing the variable that changes behaviour.
After four months, a user feedback and synthesis workshop found very high agreement between the two, alongside growing confidence and interest among users in AI-assisted interpretation as part of routine work. No agreement percentage, sample size or case count has been published.
What went wrong
The pilot also ran into problems that had nothing to do with the model.
Testing volumes came in lower than expected. Commodities were not consistently available. Devices underperformed. Connectivity was unreliable. Supervision had gaps.
Those constraints are unremarkable in community health programmes and easy to skim past, but they determine whether an AI tool works at all. A model cannot interpret a test that was never run because the kits ran out, and it cannot sync results from a phone with no signal. Uganda’s own health ministry has been building around exactly this: the offline-capable AI framework for medical supply chains unveiled in Kampala last November was designed on the premise that connectivity cannot be assumed, with its principal investigator noting that even 50MB of mobile data was enough for health workers to update the system.
Phase II is the real test
The next three months change the design. Village Health Teams will see the AI interpretation alongside their own assessment.
The question this opens is not whether the AI is accurate — Phase I addressed that. It is what happens to human judgment once the machine’s answer is visible.
Kenya provides the cautionary case. A study at Penda Health, covered by iAfrica last October, found AI significantly reduced clinical errors in history-taking, investigations, diagnosis and treatment. But more than 35% of critical safety warnings were initially ignored, and managers had to coach clinicians, track disregarded alerts and invest heavily in change management. Two patient deaths occurred during the study, both judged potentially preventable had AI alerts been followed.
That is the failure mode in one direction. The other is automation bias — the well-documented tendency to over-trust automated recommendations once a system has proved reliable, and to stop scrutinising outputs that look plausible. Both risks appear the moment the AI’s interpretation becomes visible.
The Ministry of Health, implementing districts, end users and partners have welcomed the first phase and committed to strengthening implementation as the second begins. The feedback exercise is also intended to identify improvements needed in the technology, workflows, training and support.
Why it matters beyond Uganda
Febrile illness is the most common presentation community health workers face across much of Africa, and distinguishing malaria from other causes determines whether a patient is treated correctly or sent home with the wrong medicine.
Uganda’s approach — parallel blinded testing first, visible AI second, with structured user feedback at each stage — is more methodologically careful than most African AI health deployments, which tend to launch with the tool visible from day one and measure adoption rather than agreement.
If Phase II publishes numbers, it will produce something the field is short of: evidence on whether AI-assisted interpretation actually changes what community health workers do, and whether that change helps.
