Building Uganda’s Data-Driven Health Future: An Interview With A Leading Health Data Scientist

A: A decade ago, health reports in Uganda were compiled manually from paper registers, often months late. Today, systems like DHIS2 provide near-real-time reporting from over 80% of facilities in many districts. Platforms such as UgandaEMR, OpenMRS, and the national eLMIS have digitized patient and commodity data flows.

We have moved from data scarcity to data fragmentation. The challenge now is making systems communicate. When we built the ARV stock-out prediction system in 2016, integrating LMIS transactions with DHIS2 service volumes reduced stock-out days by 30% and achieved a forecast accuracy (MAPE) below 12%. But this required custom-built pipelines. Every new partner tackling similar problems rebuilds these integrations from scratch—an enormous waste of technical capacity.

The most exciting evolution has been the shift from retrospective reporting (“what happened?”) to predictive analytics(“what will happen, and how can we prevent it?”). The ARV stock-out predictor, for example, used XGBoost with lagged features to forecast facility-level needs 7–11 weeks ahead, enabling pre-emptive redistribution. Similarly, our HIV cohort dashboard in 2019 applied Kaplan-Meier survival analysis to improve nine-month retention from 71% to 80% by identifying high-risk patient segments and underperforming facilities.

Across East Africa, the pattern is similar. Kenya leads in mobile health innovations, Tanzania is investing heavily in DHIS2, and Rwanda has achieved impressive standardization. Yet all face the same fundamental challenge: donor-driven, vertical programs that create parallel systems unable to interoperate. We are drowning in systems—but thirsting for integrated intelligence.

A: The challenges we face happen on three levels: data quality, infrastructure, and trust.

Data quality is always the toughest. When we built the HIV cohort dashboard, we had to link EMR encounters, lab results, and pharmacy refills—but they were in separate databases with inconsistent patient IDs. Without a national ID system, we had patients tracked by clinic numbers or phone numbers, none reliably unique. We spent as much time cleaning and matching data as on actual modeling. For ARV stock-out predictions, we merged weekly LMIS updates, monthly DHIS2 reports, and manual lead-time records, using tools like Prophet and XGBoost to handle missing data and make accurate forecasts.

Infrastructure is another challenge. Internet outside Kampala is often unreliable, so we design systems to work offline first, syncing to the cloud when possible. For example, the mortality review system in 2021 ran on district servers with local processing, not just in the cloud.

Trust is crucial. When the stock-out system flagged high-risk facilities, supply officers were skeptical, asking, “How does the computer know better than my experience?” We built trust through transparency—showing exactly what drove predictions, letting managers annotate forecasts with local knowledge. As the system proved accurate, trust grew naturally.

My philosophy is simple: use proven, open-source tools, focus on data quality over fancy algorithms, and design for the end user—the district health officer—not the data scientist. Gradient boosting on clean data beats deep learning on messy data every time.

A: Interoperability is one of our most critical unsolved problems, and it’s as much about governance as it is about technology. Right now, Uganda’s Ministry of Health operates a dozen major systems in parallel—DHIS2, UgandaEMR, OpenMRS, various lab systems, eLMIS, eIDSR. They rarely communicate, and when they do, it’s through brittle, custom integrations built project by project. I’ve built several myself; for example, an EMR-to-DHIS2 pipeline achieved 95% concordance with manual tallies using HAPI FHIR and Mirth Connect. But that only works for one program flow. Another partner connecting a different EMR has to rebuild everything from scratch. Clearly, this is unsustainable.

The right architectural vision comes from HL7 FHIR and OpenHIE. FHIR provides modern, RESTful APIs that align with how developers work today. It’s modular—you can start with patient demographics or immunization records and expand gradually. OpenHIE defines how health information exchange should be structured, with shared services like a patient master index, facility registry, and terminology services, all accessible via standardized APIs.

When we built the One Health platform in 2018 for zoonotic disease surveillance, we merged human case data from DHIS2, animal disease reports from agricultural databases, and environmental observations. Technically it worked—we used MongoDB for heterogeneous data and NetworkX for transmission network analysis—but it wasn’t generalizable. Today, I’d push hard for FHIR-based resource definitions for cross-sectoral reuse.

Implementation requires strong institutional structure. Uganda needs a national health information exchange—a funded entity with authority to mandate standards and provide shared services. Key steps include:

• Establishing authoritative registries as public goods, like a facility master list, health worker registry, and patient master index, all centrally maintained and accessible via APIs.

• Mandating FHIR compliance for any system procured with government or donor funds.

• Publishing Uganda-specific FHIR Implementation Guides that define our programs, identifiers, and workflows.

• Deploying an interoperability layer, like OpenHIM, for mediation, routing, and security.

• Developing health-sector-specific data governance frameworks. The Data Protection Act of 2019 provides a foundation, but we need regulations tailored to health data.

In practice, interoperability is also about collaboration. For the One Health platform, we had to negotiate data-sharing agreements between the Ministries of Health, Agriculture, and Water. That process took longer than the technical integration, but it was essential.

Finally, we need local capacity to lead this work. Makerere’s School of Public Health and Department of Computer Science could offer a dedicated program in health data analytics. Right now, expertise is mostly learned on the job. Without Ugandan FHIR specialists who understand both the standard and our health programs, interoperability will remain consultant-dependent and unsustainable.

A: Current medical curricula aren’t designed for data-intensive health systems. Statistics are often taught in isolation, with little integration into clinical reasoning. That means we’re graduating doctors who can’t interpret epidemiological curves or question why an algorithm flagged a patient as high-risk.

As AI rapidly spreads, this gap becomes dangerous. We risk a two-tier system: data scientists building models without clinical insight, and clinicians using tools without understanding their assumptions or limitations.

To fix this, we need three things. First, data literacy as a core skill for all health professionals. Every medical student, nurse, and public health officer should graduate understanding data provenance, limitations, and critical interpretation. They don’t need to be programmers, but they must ask whether a spike in cases reflects real epidemiology or just a reporting artifact.

Second, specialized health informatics training. We need professionals who combine health knowledge with technical skills in data engineering, predictive modeling, and system design. Makerere could, for example, offer a master’s in public health data science jointly between the School of Public Health and Computer Science, covering EMRs, DHIS2, FHIR, machine learning, data governance, and hands-on ministry practicum projects.

Third, continuous professional development. Curriculum reform alone isn’t enough. The Ministry should provide structured training for district health officers, HMIS coordinators, and national analysts—from basic dashboard interpretation to advanced predictive analytics and system design.

AI literacy is urgent. Tools like ChatGPT can help clinicians draft patient summaries, search literature, or explain conditions in local languages, but they can also produce plausible-sounding errors and reflect biases from Western data. Professionals must know when to trust AI, when to challenge it, and when to override it clinically. Case studies—like image recognition models failing on African patients or AI suggesting contraindicated treatments—should be part of training.

If we start now, combining data literacy, specialized training, and continuous professional development, we can ensure clinicians and data scientists work together safely and effectively in Uganda’s evolving health system.

A: Sustainability is where most digital health projects fail—and usually it’s not because of technology. One of the biggest lessons is co-ownership from the start. For the HIV cohort dashboard, we set up a technical working group from day one, including ART program managers, district focal persons, facility in-charges, M&E officers, and patient advocates. Every design decision was debated and agreed upon. It was slower upfront, but it created genuine ownership, which drove real usage.

Another key lesson is to embed systems within existing structures. Donor projects often create new “data units” outside the Ministry, which vanish when funding ends. We housed the stock-out predictor within the existing Logistics Management Unit and National ART Program, training existing staff instead of hiring external teams. Because it became part of normal operations, it outlasted the project funding.

Documentation is critical. For the mortality review system, we produced governance charters, SOPs, architecture diagrams, and job aids. When the original lead moved, the system continued running because knowledge was shared, not locked in one person’s head.

You also need to design for real users, not power users. Our initial dashboards were complex and underused. Observing district health officers, we realized they just needed to know: “Which facilities are struggling?” “Are we meeting targets?” “Where should I supervise?” We simplified dashboards to red/yellow/green indicators and priority maps—and usage soared.

Feedback loops build trust. Managers could question the stock-out predictor: “Why did it miss this crisis?” “Can we add this data source?” Failures were investigated and models updated. Users saw a learning system, not an infallible oracle.

Finally, budget security is essential. Operational costs, hosting, maintenance, and staff time must appear in Ministry budgets. Otherwise, systems die when donor funding ends.

The bigger lesson is that sustainable data systems are socio-technical. Code is only about 30% of the challenge. The other 70% is governance, stakeholder buy-in, training, documentation, budgeting, and change management.

A: The promise of AI in health is real, but it has to be deployed responsibly. Some of the most exciting applications include predictive resource allocation—we’ve already reduced stock-outs by 30% using predictive models. This approach can forecast malaria surges or other outbreaks, pre-positioning supplies before crises hit, and moving from reactive response to proactive prevention.

Clinical decision support for non-specialists is another game-changer. Uganda has just one radiologist for every two million people. AI-assisted chest X-rays for TB or AI-analyzed retinal photos for diabetic retinopathy can extend specialist expertise to local health centers. The AI doesn’t replace judgment—it augments it, helping general clinicians reduce errors.

AI can also improve supply chain visibility and diagnostic accuracy by integrating data across facilities, labs, and patient history. Even large language models like ChatGPT have potential—helping health workers draft patient notes, query treatment protocols in local languages, or summarize complex medical literature. AI tutors could also simulate clinical cases for students.

But there are risks. ChatGPT can confidently generate incorrect information and lacks context specific to Uganda. Health professionals need training to know when to trust AI, when to challenge it, and when to override it clinically.

Responsible deployment requires strong data governance, clear protocols for ownership, consent, and de-identification; algorithmic transparency, with training data, accuracy metrics, and limitations clearly documented; local validation to ensure models work for Ugandan populations; local capacity building, so Ugandan data scientists can develop and manage AI tools; and equity audits, to ensure models serve all populations fairly, including refugees and marginalized communities.

Finally, AI is just one tool. It can’t replace investments in primary healthcare, clinician training, or infrastructure. Even the most sophisticated model is useless if districts lack the resources to act on its recommendations.

Q7: What is the top policy priority to accelerate Uganda’s data-driven health transformation?

A: Uganda needs a National Health Informatics Institute—a funded, authoritative body that handles standards, interoperability infrastructure, capacity building, and technical assistance.

Right now, our digital health landscape is fragmented. Donor projects work in isolation—PEPFAR funds HIV EMRs, the Global Fund supports DHIS2, CDC invests in lab systems, UNICEF backs vaccine logistics. Each makes sense on its own, but together they create dysfunction: data can’t flow, and we keep reinventing the same solutions.

As someone building integrations, I see this waste firsthand. Every partner building EMR-to-DHIS2 pipelines solves a problem that should be solved once and made available to everyone. We lack the connective tissue to make these systems work coherently.

The Institute would change that. It would set and enforce standards, publishing Uganda-specific FHIR Implementation Guides and maintaining terminology mappings like ICD-10, RxNorm, and LOINC adapted to our context. It would run shared interoperability infrastructure—a patient master index, facility and health worker registries, terminology services, and an interoperability layer like OpenHIM for secure data exchange.

It would provide technical assistance, with health informaticians and FHIR specialists helping districts and partners implement dashboards, integrate systems, and troubleshoot issues. And it would offer training and certification, from basic DHIS2 configuration to advanced analytics and system architecture, creating recognized career paths for health informaticians.

The idea is maximum leverage. Training data scientists or funding analytics tools is wasted if systems can’t interoperate. The Institute acts as a force multiplier, making individual investments reinforce each other rather than creating silos.