PDX: Africa's AI-assisted health early-warning system must also be cyber-resilient

A political call to action in New York

On 25 September 2026, African leaders, health ministers and international partners adopted the New York Call to Action on Cross-Border Preparedness Intelligence in Africa. The announcement published the following day by the WHO Regional Office for Africa calls for stronger cross-border cooperation and responsible use of artificial intelligence to anticipate infectious diseases capable of causing public-health emergencies.

The meeting highlighted the Preparedness Data Exchange (PDX). Launched by WHO Africa in March 2026, the intelligence system is intended to combine several categories of information: risk assessments, International Health Regulations capacities, primary-care readiness, climate intelligence, workforce data, laboratory trends, emergency operations and media monitoring. Its purpose is to help officials detect danger earlier and make evidence-based decisions across fragmented sources.

That is a significant opportunity in a region where outbreaks, population movement and climate events can cross administrative borders quickly. Connecting health intelligence from 47 countries and adding AI capabilities, however, also makes PDX a sensitive digital infrastructure. The reviewed announcements do not report a cyberattack against PDX. The security issue is preventive: safeguards need to exist before operational dependence makes them difficult to add.

How AI changes the risk model

A conventional dashboard presents data. An AI-assisted system may also rank alerts, connect events and recommend where attention should go. That creates value while expanding what must be protected. A false or decontextualised record may no longer remain an isolated error; it can influence a score, priority or summary delivered to a decision-maker.

The first risk is therefore integrity. Compromised accounts, an exposed API or a manipulated partner source could introduce misleading observations. In health surveillance, an attacker does not need to take over the whole platform. It may be enough to inflate one signal, conceal a genuine event or generate enough noise to slow analysts. Controls need to establish the origin, timestamp, quality and modification history of consequential inputs.

The second risk is confidentiality. Aggregated analysis can appear anonymous, yet a precise location combined with a rare condition and a small population may permit re-identification. Cross-border cooperation should not mean indiscriminate replication of patient-level records. The safer principle is to share the minimum required information, using aggregation, pseudonymisation and retention rules tailored to the public-health purpose.

The third risk is availability. An early-warning platform used during an outbreak becomes attractive for extortion, sabotage and disinformation. It needs to tolerate weak connectivity, retain manual procedures and allow national teams to continue working when a central service or cloud provider is unavailable.

Trust cannot rest on a score

WHO has stated that AI-enabled preparedness is not about replacing epidemiologists or public-health leaders. That boundary should become a technical property rather than a policy slogan. An alert should expose the data that triggered it, the transformations applied, uncertainty and the person responsible for validating action. Without that traceability, a plausible but wrong recommendation may be accepted because its interface looks authoritative.

Systems should distinguish three layers: observed data, model inference and human decision. Audit records must reconstruct that path without unnecessarily reproducing personal information. Models and analytical rules should be versioned so that an update cannot silently change how the same evidence is interpreted.

The same discipline applies to suppliers. An AI or cloud contract does not by itself establish accountability. Authorities need to know where data is processed, which subcontractors can access it, whether information can be used to train other models, how long backups survive and how data can be recovered in a usable format.

Five priority controls for PDX and similar platforms

1. Federated identity and least privilege. Each country and partner institution should receive constrained roles, phishing-resistant multifactor authentication and rapid revocation procedures.
2. Verifiable provenance. Critical feeds should be signed or authenticated, timestamped and linked to an identified source. Corrections must leave an audit trail.
3. Validation before automation. High-impact recommendations should require human review and display contradictory evidence, not only a model conclusion.
4. Regional resilience. Restorable backups, disconnected operating modes and joint exercises should cover cloud failure, partner compromise and data corruption.
5. Independent testing. Assessments should combine application security, access control, model robustness, bias, privacy and health-disinformation scenarios.

Health sovereignty is part of the design

The New York statement says countries should retain control of their data while enabling the intelligence sharing required for collective preparedness. That is the right objective, but it needs a concrete architecture. Sovereignty is not only the location of a server. It includes the capacity to understand models, enforce access rules, audit processing, change providers and maintain operations without irreversible dependence.

The African Union's Continental AI Strategy already provides a policy foundation built around inclusion, rights, safety and adaptation to African needs. PDX now offers a practical case for translating those principles into shared technical requirements. A continental approach can also prevent each health ministry from negotiating alone over the same controls and supplier risks.

What decision-makers should retain

AI can improve preparedness by connecting signals that isolated teams might recognise too late. Speed does not compensate for altered data, an unexplainable decision or an unavailable platform during a crisis. PDX should therefore be measured not only by the quality of its alerts, but also by whether countries can challenge an output, trace its evidence and continue operating when technology fails.

Africa has an opportunity to build AI-assisted public-health infrastructure with security from the beginning. Deferring these controls until after widespread deployment would create technical and institutional debt precisely when the platform becomes most important.

— ◆ —