What Had to Exist Before AI Could Matter in African Healthcare

By Tosin Clegg

Before clinical intelligence could become useful, the harder problem was building the data foundations that healthcare systems across Africa still lack. For Tosin Felixson-Yusuf, an Engineering Expert he describes a pivotal moment where this problem became real for him. According to Tosin, “The first time the gap became real for me was not in a policy document or a conference panel. It was inside the data itself. We were working on building longitudinal patient records, the kind of continuous clinical history that follows a person across visits, across facilities, across time. The principle is straightforward. If you visit a hospital today and return six months later, your record should know you. If you move to a different facility, your history should come with you.”

“In practice, it often does not. A different hospital frequently means a new record. A change in EMR vendor, which happens more often than
people outside the industry realise, can mean the migration does not carry existing records across at all. The data schemas differ. The work of moving records between systems that were never designed to speak to each other is significant. A new system starts fresh and years of clinical history effectively disappear.” He added.

For Felixson-Yusuf this wasn’t just unique to Nigeria. It happens across the world. But in countries where a national patient identifier exists, where a central reference point connects a person to their records regardless of which system holds them, the damage is contained. Here it compounds. And what compounds with it is something larger than a records problem. The absence of longitudinal data is also the absence of the foundation that intelligence in healthcare is built on.

At some point the gap became structural and Felixson-Yusuf had to step in. Recollecting his approach to this he noted, “I was building this at Helium Health, where we had grown to become one of the largest EMR providers in sub-Saharan Africa and one of the widest reaches of patient records in the region. The scale made the gap more visible, not less. The more records you hold, the more clearly you see what is missing between them.”

He also pointed out that away from the engineering problem and the scale of what it represents becomes clearer. Fewer than 18% of Nigerian hospitals currently use electronic medical records. That number sounds modest until you understand what it actually means in practice. The 18% is not evenly distributed. It clusters in urban centres, in private tertiary and secondary facilities, in the parts of the healthcare system that serve the minority of Nigerians who can afford private care. The primary health facilities where most Nigerians first encounter the health system remain largely paper-based.

The result is that the patients most likely to appear in any digital health dataset are the least representative of Nigeria’s actual disease burden, demographic spread, and clinical reality. Any model built on that data inherits that skew structurally. It is not a data problem in the technical sense. It is a representation problem with direct consequences for what intelligence built on that data can and cannot do.

The 18% also overstates what is practically usable. That figure represents data dispersed across dozens of separate platforms with no common architecture, no shared standards, and no mechanism for exchange between them. A patient record in one system is invisible to every other. The continuity that makes data medically meaningful does not exist at the system level. It has to be reconstructed through engineering work that should not have to exist.

The problems compound further when you get inside the data itself, Felixson-Yusuf discloses and add that, “When children are registered at clinical facilities by their parents, most platforms have no mechanism to transition that dependency when they become adults. The record continues to reference a guardian rather than an individual, quietly introducing a category of error that accumulates across millions of entries over years.”

“Consent introduces another layer of complexity that is rarely discussed openly. To legally manage patient records on a digital platform, verifiable consent must be obtained from each patient. That process is more difficult to implement correctly than it sounds. It requires specific legal architecture, documented consent workflows, and
ongoing compliance management. Without it, digitally managing patient records without consent is a breach regardless of the technical quality of the system.”

The Engineering Expert also pointed out that In healthcare, security is not simply an IT consideration but more of a question of national sensitivity. Health data is among the most sensitive information any government holds about its citizens. Globally, healthcare has become the most targeted sector for cyberattacks. In 2024 alone, 276 million patient records were exposed across reported breaches, at an average cost of $7.42 million per incident, the highest of any
industry. Building healthcare data systems in this environment requires a standard of engineering intentionality that goes well beyond what most technology frameworks demand. Encryption across the entire workflow. Access controls that expose real patient data only to those with a direct clinical reason to see it, the treating clinician, the patient
themselves, and not to engineers, analysts, or anyone else. The security architecture of a healthcare system cannot be retrofitted. It has to be foundational. The engineering decisions made at the beginning determine what is possible later.

The contrast between where Africa is and where it is heading becomes clearer when you look at what national mandate actually produces, Felixson-Yusuf explains. He further argues that, “Rwanda is the most instructive example on the continent. After systematically implementing a national EHR across all public health facilities, Rwanda launched a National Health Intelligence Centre in April 2025, a centralized platform that uses artificial intelligence and predictive mathematical models to anticipate disease outbreaks up to three months before they occur. Instead of spraying entire regions to combat malaria, Rwanda’s Ministry of Health can now target only the high-risk zones its
models identify. That precision is not the product of superior technology. It is the product of years of deliberate data infrastructure building that created a foundation clean and connected enough for intelligence to work on top of it.”

“Kenya is moving in the same direction. Its Digital Health Act of 2023 established legal frameworks for data protection, interoperability standards, and the regulation of digital health services, building the governance architecture that has to exist before data can move freely and safely across a health system. Nigeria is making its own deliberate push. The Nigeria Digital in Health Initiative, launched in March 2024 and endorsed by the National Council on Health in November 2024, represents the most ambitious attempt yet to unify Nigeria’s fragmented digital health landscape. Its three pillars, an interoperable digital health services network, a health claims exchange, and a health information exchange, are designed to produce exactly what has been missing: a national data space where clinical information can move securely across facilities, states, and both public and private systems.”

One of its most important and under appreciated engineering decisions is the mandate for offline access. Most of the facilities that remain undigitised are undigitised precisely because they cannot guarantee reliable power or internet connectivity. A system that requires connectivity to function excludes the very demographics it is meant
to serve. Offline-first architecture, where data is captured locally and synchronised when connectivity becomes available, bridges that gap. It makes mobile-based and even USSD-based data entry possible, reaching clinicians and facilities at the bottom of the infrastructure pyramid. That is the right engineering instinct.

Across the continent, a parallel infrastructure has been quietly doing foundational work for years. DHIS2, the District Health Information Software developed by the University of Oslo and deployed in more than 70 countries at national scale, covers health data for over 40% of the world’s population.

Back inside the work at Helium Health, as the infrastructure scaled, the nature of the problem shifted as he further describes that, “Having records is not the same as having intelligence. The systems we had built were OL TP systems, transactional by design, optimised for capturing what happened in a clinical encounter in real time. To move from digitisation to analysis, from recording to understanding, we needed to build the analytical layer on top: OLAP infrastructure that could aggregate, model, and provision intelligence across the entire patient corpus. That is a fundamentally different engineering problem. One is about capturing events accurately. The other is about asking what those events mean at scale.
The first challenge was identity. T o build longitudinal intelligence, to model a patient’s risk profile, trace the progression of a condition across time, understand how a population responds to treatment, you need to know that the records you are looking at belong to the same person. In the United States, a Social Security Number is a
prerequisite for accessing healthcare. In Nigeria, we have the National Identification Number and the Bank Verification Number, but neither is required to receive treatment. A patient can walk into a facility and be registered without any national identifier at all. The result is that the same person can exist as multiple separate records across the system, registered differently at different facilities, across different visits, across different EMR vendors, with nothing connecting them.”

“We had to build a bespoke identity model to solve this. Using multiple indirect signals traced across patient activity and modelled against each other, we built a system that could determine with sufficient accuracy that an individual we had seen before was the same person we were looking at now. Not perfectly. But with enough confidence to begin constructing a longitudinal view. The constraint that made this harder than it sounds is the same constraint that makes it correct. Our security
architecture de-identifies patient records before they enter the analytical pipeline. The transactional system holds the full clinical record with identifiers, because a clinician treating a patient needs to know who they are treating.”

But when that data moves into the analytical environment for modelling, the direct identifiers are stripped. The pipeline sees signals, patterns, encoded references, not names, not contact details, not anything that could directly reveal who a patient is. This is the right way to build it. It follows the principle of least privilege. It aligns with where health data regulation is moving across the continent. But it means the identity reconstruction problem has to be solved without the most obvious signals. The model has to approximate who a person is using only what it can observe indirectly. That is exactly what privacy-preserving analytics is supposed to demand of you. The constraint is not a limitation. It is evidence that the architecture is working the way it should.

Felixson-Yusuf now establishes that the next phase of African healthcare is not difficult to imagine if you have been inside the infrastructure long enough to see where it is heading. When the investments in health information exchange mature, when interoperability becomes operational rather than aspirational, a patient will be able to walk into a hospital in Lagos, get referred to a specialist in Abuja, and arrive with their full clinical context already there, their history, their medications, their previous diagnoses, carried not in a paper folder but in a connected system that knows
who they are.

That operational reality unlocks everything else. Disease surveillance becomes automatic rather than manual. Epidemiological signals that currently take months to surface will emerge in real time. Rwanda is already doing this, predicting outbreak hotspots three months in advance using AI models built on a national data foundation that took years of deliberate infrastructure work to create. Nigeria is building toward the same capability. So is Kenya. So is Ghana. The continent is moving and the direction is clear.

But the operational layer is only part of the picture. The intelligence layer requires something more specific, the right datasets, curated and standardised with enough discipline to generate real world signals that actually reflect African clinical reality. Not models trained on data from populations that do not share our disease burden, our comorbidities, our treatment pathways, or our environmental context. Models built on our data, for our context, validated against our outcomes. That work is harder than importing a model from elsewhere and calling
it done. It requires engineering discipline at the data layer before a single model is trained. But when it is done correctly, what it produces is not a pale imitation of what the Global North built. It is something more valuable, clinical intelligence that actually works in the conditions where it will be used.

Healthcare is at a similar inflection point. The burden of disease is high. The ratio of doctors to patients is acute. The infrastructure gaps are real. But those same constraints are what make clinical decision support systems more valuable here than anywhere else on earth. A system that helps one doctor serve thousands of patients better is not a productivity tool in Africa. It is a fundamental shift in what is possible. The raw materials are here. The workflow is unique. The market is enormous and underserved. Investment is beginning to follow, and the returns, both humanitarian and commercial, will be significant for those who understand what they are investing in early enough.

In concluding, Felixson-Yusuf shares that, “What I am building toward is the moment people look back and realise that Africa’s healthcare constraints were not a disadvantage. They were the reason we built something the world had not seen before.”