AI in African healthcare: what 18 months of published deployments actually show

Executive summary

The case for AI in African healthcare is usually told in the wrong direction. Most of the public discussion starts from algorithms and proposes infrastructure as a follow-on problem. The deployments that have actually run for years tell a different story. They are bottlenecked by data quality and pipeline discipline, not by model architecture. They are gated by clinician trust, not by communication strategy. They were forced by power and connectivity constraints to choose architectures that no peer-reviewed paper would have recommended on first principles. They operate inside data-protection regimes that vary materially by jurisdiction, which means a multi-country deployment is almost always a portfolio of separately governed pipelines rather than a single platform.

This whitepaper reads the published African healthcare AI literature from 2019 to 2024 alongside our practitioner observations and lays out five lessons that hold up across both. We anchor those lessons in deployments that exist in the public record: Babyl Rwanda's primary-care partnership ^[4], Zipline's drone-delivery integration with national supply chains ^[8], Vula Mobile's specialist-referral application across the SADC corridor ^[11], MinoHealth AI in Ghana ^[16], Ilara Health's clinic-network model in Kenya ^[12], and the qXR and CAD4TB tuberculosis screening tools used under WHO's 2021 conditional recommendation ^[7]. We are explicit about what the literature supports, where it stops, and what we have observed independently. We do not name client deployments, and we do not extrapolate from our own engagements to claims about the continent. The published evidence remains concentrated in a small number of countries and institutions, which is itself a finding worth taking seriously.

01 /Data quality is the operational bottleneck

Owoyemi and colleagues' 2020 review of AI for healthcare in Africa was unambiguous on this point ^[1]. The constraint on AI in African health systems is not modelling. It is the data substrate the modelling assumes. The same argument runs through Schwalbe and Wahl's Lancet piece on AI and global health, which framed the operational case as a question of data infrastructure and governance maturity ^[2]. The IFC's 2023 portfolio observations on AI in emerging-market healthcare lands in the same place from a development-finance perspective ^[10].

In practical terms, the universal background is duplicate patient records, inconsistent diagnostic coding, free-text fields where structured fields existed, missing demographic data, and timestamp drift between systems. Every healthcare AI engagement we have done in this region has been bottlenecked by data quality, not by modelling. We spend more engineering time on entity resolution and pipeline integrity than on model development. The conference talks understate it. The published literature increasingly does not.

The implication is uncomfortable for procurement-driven programmes. A model that performs well on a curated retrospective dataset will degrade against a live data stream that includes free-text history fields, partial demographic information, and clinically meaningful but unstructured clinician annotations. The fix is upstream of the model. It is data infrastructure that should have existed before AI was on the agenda, and the build cost of that infrastructure is often the largest line item on a deployment that is sold as an AI project.

There is a budgetary corollary worth naming. If a delivery plan allocates twenty percent of the build cost to data infrastructure and eighty percent to modelling, that plan has misread the operating environment. Our experience, consistent with the IFC's portfolio framing ^[10], is that the ratio in a deployment that survives contact with live operations sits closer to the inverse. Identity resolution alone, the apparently boring problem of deciding whether two records describe the same patient, can dominate the first six months of a programme. Hospitals in our experience routinely carry duplicate-rate estimates between fifteen and forty percent in legacy systems, and the AI tools downstream perform very differently against deduplicated and non-deduplicated streams. The choice of whether to invest in entity resolution before model development is, in a real sense, the most consequential modelling decision the programme will make.

A second corollary is the labelling question. Many published African healthcare AI papers rely on retrospective labelled datasets that were curated specifically for the study. Live operations rarely have that luxury. Diagnostic codes are inconsistent across clinical sites and even between clinicians within a site, free-text fields contain clinically rich information that a structured-field-only pipeline will miss, and the cost of generating high-quality labels at production scale is rarely accounted for in the original business case. The teams that have made this work have generally invested in a structured clinician-feedback loop, in which model outputs are reviewed and corrected by clinical staff in a disciplined way, and the corrections feed back into the next retraining cycle. That is a workflow problem, not a modelling problem, and it requires institutional discipline that is difficult to procure.

02 /Clinician trust is earned through structured validation, not communication

The cleanest illustration in the public record is the contrast between Babyl Rwanda and Vula Mobile. Babyl Rwanda paired Babylon's symptom-checker AI with the country's primary-care system and reached over two million registered users at peak ^[4]. The parent company's commercial collapse in 2023 ^[3] complicates the legacy, but the deployment did establish that a chatbot triage front-end can scale in a primary-care system. Reach is not adoption. The question of whether the AI's recommendations were trusted by clinicians, and whether they durably shifted clinical decisions, is harder to answer from public sources.

Vula Mobile took a quieter approach. The South African application connects general practitioners to specialists for asynchronous consult, using structured templates and image attachments. It has expanded across the SADC region and is now used by a workforce the operator self-reports in the tens of thousands ^[11]. The model is incremental: it does not replace the clinician's judgment, it routes harder cases to the right specialist faster. That framing made the trust question tractable.

In our practitioner experience, the deployments that durably shifted clinical practice all involved long pre-deployment validation cycles, named local champions, and explicit scoping of the model's role. Adoption changes when clinical leadership has seen the model being honest about its disagreements, not when an executive sponsor circulates a deck. We have rarely budgeted enough time for this work in the first iteration of a delivery plan. We do now, and the published deployments that have outlasted their funding rounds suggest this is industry-wide.

There is a specific organisational pattern worth describing, because it recurs. The model is presented to the clinical leadership in a structured side-by-side review against historical cases, with the chief clinical officer or department head reviewing each disagreement with the data-science team. The review surfaces the cases where the model was right, the cases where the clinician was right, and the cases that are genuinely ambiguous. The proportions matter less than the existence of the conversation. Once the clinical leadership has formed its own view of where the model is reliable and where it is not, the floor staff follow. Without that step, a triage or referral support model is treated as a black box that nobody is willing to trust under time pressure, and adoption stalls in a way that cannot be unblocked by a communications campaign.

A second pattern from the published Babyl experience is worth flagging carefully. Reach metrics (registered users, sessions delivered, queries handled) are easy to publish and operationally meaningless. They do not tell you whether clinicians changed what they recommended, whether patients changed what they did, or whether downstream outcomes improved. The Babyl Rwanda partnership produced impressive reach numbers and an unclear clinical-outcome record ^[4], and the parent company's broader trajectory ^[3] complicates any retrospective view. The lesson is not that the project failed, because for parts of the system it clearly worked. The lesson is that the metrics that survive scrutiny are the ones tied to clinical decisions and patient outcomes, and the metrics that look good in a press release are usually neither.

03 /Infrastructure shapes architecture more than the modelling problem does

Mollura and colleagues argued in 2020 that AI deployment in low and middle income country radiology requires co-designed workflows rather than transplanted reference architectures ^[15]. The argument has aged well, and the deployments that scaled bear it out.

The radiology AI tools with the most durable production track record on the continent (qXR from Qure.ai, CAD4TB from Delft Imaging) were designed for offline-friendly inference on commodity hardware. Multi-site validation including African data preceded the WHO's 2021 conditional recommendation that computer-aided detection software may substitute for human readers in adult tuberculosis screening ^[6][7]. That recommendation has materially changed TB screening programmes, and it would not have been operationally feasible if the underlying tools required hospital-grade GPU infrastructure or persistent connectivity.

The pattern repeats outside radiology. Zipline's drone delivery network in Rwanda, Ghana, Nigeria, and Cote d'Ivoire integrates routing and demand forecasting models into the national medical supply chain ^[8]. The architecture concedes to the operating environment. Drone hubs are positioned for the road and electricity reality on the ground, not for an idealised topology. Power reliability shapes which models can be served and where. We have rebuilt deployments to function in degraded mode during outages, which usually means a smaller local model that runs on a single GPU node with battery backup, rather than the cloud-served reference architecture the original brief assumed.

The lesson is not that the published cloud-first reference architectures are wrong. It is that they assume an infrastructure baseline that is not the median in the operating environment, and the deployments that have lasted made the architectural concession explicitly.

There is a related point about model drift in this context. A patient-flow forecasting or admissions model trained on three years of historical data assumes the operational environment of the next twelve months looks like the operational environment of the last thirty-six. In African health systems, that assumption fails routinely. A hospital changes its referral protocol with a regional facility, a national insurance scheme rolls out, a new clinic opens nearby, and the historical pattern stops applying. The architectural question is not just whether the model can run with intermittent connectivity. It is whether the retraining cadence and monitoring discipline are built in from the beginning. We have seen forecasting models perform well on backtests, deploy cleanly, and fail within a matter of weeks because the underlying data distribution shifted. The fix is not a better model. It is faster retraining, tighter monitoring, and an operational culture that treats model degradation as expected rather than exceptional.

04 /Consent variance is a design constraint, not a legal overhead

A multi-country African healthcare AI programme is almost always a portfolio of separately governed pipelines rather than a single platform. Nigeria's Data Protection Act 2023 ^[14] is the current statutory frame for sensitive personal data in clinical contexts in Nigeria, replacing the NDPR 2019 with stronger lawful-basis requirements and a clearer role for the Nigeria Data Protection Commission. South Africa's POPIA, in force since July 2020 ^[13], has its own treatment of cross-border data movement and special-category information.

These are not interchangeable. A retraining workflow that is straightforward under one regime is sensitive under the other. In our practice, we negotiate separate data use agreements with each network and routinely end up with different permissions on retraining, different audit windows, and different obligations on cross-border data transfer. The result is that what looks externally like one programme operates internally as several parallel data lifecycle pipelines. This is not efficient. It is what a careful deployment looks like in this region.

The WHO Africa regional strategy on digital health acknowledges the same gap between national-level interoperability ambitions and the implementation reality on the ground ^[5]. The published case studies that have continued to run have generally taken the consent and governance question seriously enough to design around it from the beginning, rather than treating it as an artefact that the legal team will sort out before launch.

There is a deeper question buried inside the consent regime. Most healthcare consent forms in use across the region were drafted before the question of model retraining on prospective data was a live operational issue. They cover research, audit, and clinical care, often broadly. They do not generally address the question of whether a model trained partly on a patient's data may be deployed at a different institution, or whether a derivative model trained on aggregated data may be commercialised. The literature on this gap is still emerging. The IFC's report acknowledges it indirectly ^[10], and Schwalbe and Wahl flagged it in 2020 ^[2]. We approach this in practice by negotiating retraining and cross-institutional deployment terms separately from the underlying clinical-care consent, and by building audit logs that make the data lineage of any deployed model legible to the data-protection authorities in each jurisdiction. This is administrative work that does not appear on a model-performance chart, but it is what determines whether the deployment can survive a regulatory review two years after launch.

05 /Capability over procurement

Brookings argued in 2023 that the African AI opportunity in healthcare runs through institutional capability building rather than vendor selection ^[9]. We agree, and the longer-running case studies bear it out.

Domain-focused groups like MinoHealth AI in Ghana have applied machine learning to specific clinical bottlenecks (radiology shortages, in MinoHealth's case) by building close working relationships with the clinical groups whose work the tool affects ^[16]. Clinic-network operators like Ilara Health in Kenya have bundled diagnostics with software in primary-care clinics, with the software treated as a workflow and operations problem rather than a standalone product ^[12]. The institutions that have made these models work are not unusually clever at modelling. They are unusually disciplined about workflow integration, retraining cadence, and the slow work of clinician engagement.

The institutions that bought a model and called it strategy mostly do not feature in the longer-running case studies because the deployments stopped working. We do not have a clean public record on the failures, which is itself a feature of the literature: vendors and operators publish on what worked and remain quiet on what did not. The procurement-led failure mode is over-represented in our private experience and under-represented in the public archive.

The capability frame implies a different procurement question. Not which vendor has the strongest demo, but which delivery configuration (vendor, internal team, clinical leadership, regulatory navigator) the institution can sustain across a retraining cycle and a leadership transition. The published deployments that scaled all answered that question before they answered the modelling one.

06 /What the literature does not yet cover

Three gaps are worth flagging, because they shape how readers should interpret everything above.

First, the published evidence is concentrated in a small number of countries (Rwanda, South Africa, Kenya, Ghana, Nigeria) and a small number of institutions within those countries. The experience of secondary and tertiary facilities outside the well-funded teaching hospitals and clinic networks is largely missing from the literature. We do not know with confidence whether the lessons above hold for smaller facilities without the technical infrastructure of the published cases.

Second, longitudinal data is thin. Most published deployments report on the first 12 to 24 months. Operationally, AI in healthcare is a multi-year discipline, and the questions about retraining cadence, model drift, and the durability of clinical adoption need to be answered over five-plus years. The literature is not yet old enough to settle them.

Third, failure cases are under-published. We have observed deployments that quietly stopped working when an executive sponsor moved on, when a vendor's funding round dried up, or when a clinical service line was reorganised. These do not generally get written up. The result is a public record skewed toward survival bias, which makes the published lessons more useful as a positive checklist than as a complete picture of the operating environment.

07 /What to take from this

Eighteen months of published material and our own practitioner experience converge on the same direction: AI in African healthcare works when it is treated as a capability programme that takes data infrastructure, clinician trust, operational architecture, and jurisdiction-by-jurisdiction consent design seriously from the beginning. The deployments that did not survive that test are mostly absent from the public record because their operators did not write about them. The lessons are useful as a positive checklist, with the caveat that they do not describe the failure modes the literature does not yet capture.

Coderex advises ministries of health, hospital systems, and AI vendors on the operational design of healthcare AI programmes for African contexts: data infrastructure investment first, structured clinician trust cycles with named clinical leadership, architectures designed for the power and connectivity reality of the deployment site, and consent and data-protection design that respects the jurisdictions actually in scope. The work is more administrative than algorithmic. The institutions that succeed treat that as a feature.

Expect the WHO conditional recommendation on TB screening to be extended to at least one additional clinical use case before 2028 as the multi-site validation literature matures. Expect the first published longitudinal study of a continental healthcare AI deployment past the five-year mark to land in 2027 or 2028, repricing how durably any of these systems shift clinical outcomes. Expect at least one African data-protection authority (Nigeria's NDPC, South Africa's Information Regulator, or Kenya's ODPC) to issue a formal finding on retraining or cross-institutional model deployment within the next 18 months, which will reshape how multi-country programmes are structured. Expect the failure-case literature to start filling in over the next two to three years as funders and academic groups begin commissioning explicit retrospectives on deployments that did not last.

Methodology

This whitepaper is a structured reading of African healthcare AI literature published from 2019 to 2024, cross-referenced with our practitioner observations from health-sector engagements in West, East, and Southern Africa. Sources include peer-reviewed journals (Lancet, Frontiers in Digital Health, Radiology, Nature Scientific Reports), regulator publications (WHO, NDPC Nigeria, Information Regulator South Africa), and reporting in Reuters, Financial Times, Quartz Africa, Rest of World, Associated Press, and TechCrunch. Where we cite vendor or operator self-reported figures (Babyl Rwanda registered users, Vula Mobile workforce reach), this is flagged in the references.

We have not run a primary survey for this whitepaper, and we do not name client deployments. We treat our practitioner experience as one input alongside the literature, not as a separate evidentiary base. Where the published material and our experience diverge, we say so. Where they converge, we treat the convergence as itself useful evidence, while noting that we are not a representative sample of African healthcare AI delivery practice and our exposure is biased toward larger institutions and longer engagements.

The literature is concentrated, longitudinal data is thin, and failure cases are under-published. Readers should weight any conclusions accordingly. We will revisit this with what we hope is a wider evidence base in 2027.