From Maps to Action: Reflections on Our Tamatisha TB Dissemination Workshop in Nairobi

Last week, our team travelled to Nairobi for one of the most rewarding moments in any project's lifecycle: the point where the work leaves the room and enters real-life. Over two days, we sat down with National Tuberculosis Programme (NTP) leadership, our partners at the Centre for Health Solutions (CHS), and a wider group of implementing partners to share what the Kenya Tamatisha project has produced.

I was there alongside Bernadette Nkhata, our Public Health Project Manager, and Matthys Potgieter, our CTO. Between the three of us we covered the why, the what, and the how, but the real energy in the room came from the people who are actually using these tools in the field.

Why we were there

Tamatisha is about a simple, stubborn problem. Tuberculosis remains one of Kenya's leading infectious disease challenges, and the resources to fight it, screening teams, clinical staff, treatment support, are finite. The question isn't only how much effort to apply, but where to apply it for the greatest effect.

Our job over the past several months has been to turn Kenya's own programme data into decision support: AI and spatial models that help teams move from broad geographic coverage to precision-targeted outreach. The dissemination session was our chance to translate technical outputs into something programme managers and county teams could pick up and run with.

Three findings, three questions

We structured the first day around three questions that TB programmes ask constantly.

Where is TB risk likely highest? Our AI TB Risk Model predicts, at sub-ward level across Kenya, the proportion of individuals likely to test positive for TB. It does this by combining active case finding (ACF) screening data from Kenya's TIBU system, data on population, socioeconomic, spatial, and health-system variables. The output is a ranked list, the top 30% of wards by predicted positivity, plus specific high-density locations within priority sub-wards for community screening. The message we wanted teams to leave with was practical: this is a starting point for planning, not a substitute for field feasibility checks, and results should feed back to sharpen the model over time.

Where are the gaps in TB prevention? Our second finding looked at the TB preventive therapy (TPT) cascade, from identifying contacts of index patients, to evaluating them, to starting treatment, to completing it. Mapping the cascade revealed something the room recognised immediately: contact identification is often strongest in the most deprived and remote areas, but the drop-off begins after clinical engagement. The bottleneck isn't finding contacts anymore, it's converting them into treatment.

Who is at greatest risk of poor outcomes? Our third model predicts drug-susceptible TB treatment success and unfavourable outcomes (failure, loss to follow-up, death) at ward level. A few patterns stood out. Poor outcomes cluster in densely populated areas, sometimes hidden inside otherwise well-resourced settings. Malnutrition emerged as the single strongest correlator of poor outcomes. And interestingly, while HIV-positive individuals face higher clinical risk, areas with higher HIV burden actually showed better treatment success, a likely testament to Kenya's strong TB/HIV service integration. We mapped where high vulnerability intersects with low treatment success, so those areas can be prioritized directly.

Under the hood: the technical days

Presenting findings is one thing; handing over the machinery is another. That's why the bulk of the workshop was dedicated to technical capacity building — and this is where Matthys and the technical team went deep.

We walked through the EPCON data ecosystem end to end: the platform architecture, the secure TIBU API connection, and an aggregation approach we care a great deal about, ward-level precision with no patient-level data transferred, refreshed on a 24-hour cycle. We were candid about model inputs and their limitations, from population and socioeconomic layers to spatial data and programme data, including the genuine challenge of reconciling facility-linked versus address-linked notifications.

From there we opened up the methodology: how the outcome variables are defined, how features are selected, how models are trained and validated, and how the top 30% priority wards are chosen and what that threshold does and doesn't mean. We spent time on appropriate use: what "predicted positivity rate" means, where facility-linkage bias creeps in, and the common misinterpretations to avoid. A model is only as useful as the trust and understanding behind it, so we would rather teams know its limits than over-read its outputs.

The conversation we came for

The moment I valued most wasn't a slide, it was the discussion. Does the model output match how ACF teams actually choose sites today? What additional data would improve accuracy, and what refinements are needed before outputs are used operationally? These are exactly the questions that turn a good model into an adopted one, and the county teams brought hard-won operational realism to every one of them.

Where we go next

Dissemination isn't an endpoint. As we move on, our focus is on aligning these models more tightly with NTP data structures and workflows, closing the feedback loop between field verification and model refinement, and supporting teams as they put prioritization outputs to work.

My thanks to the NTP, the Ministry of Health, the Centre for Health Solutions, and every county officer and partner who gave us two days of sharp questions and generous engagement. The maps are made.

Now comes the part that matters: using them.