AI Diagnostic Accuracy in Sub-Saharan Africa: A 2025 Field Study

The deployment of AI-assisted diagnostics in high-income healthcare systems has been extensively studied. Far less attention has been paid to how these systems perform in low-resource environments — where patient volumes are high, clinician capacity is constrained, and the cost of misdiagnosis is acute. This study addresses that gap.

Between January and December 2025, Be Okay's AI triage engine was deployed across 47 primary care clinics in Ghana, Rwanda, and Nigeria. The engine analyzed patient-reported symptoms, vital signs captured via low-cost Bluetooth devices, and structured intake questionnaires to generate a prioritized clinical assessment and suggested diagnosis for each consultation.

A total of 84,312 consultations were analyzed. In each case, the AI assessment was compared against the final clinician diagnosis, which served as the ground truth. The AI system achieved an overall diagnostic accuracy of 81.4% — meaning that in more than four out of five consultations, the AI's primary or secondary suggested diagnosis matched the clinician's final assessment.

Performance varied by condition. For malaria — the most common presenting complaint across all three countries — the AI achieved 89.2% sensitivity and 84.7% specificity. For hypertension screening, sensitivity reached 93.1%. Performance was lowest for complex multi-morbidity presentations, where clinician expertise remained superior.

Critically, the AI system demonstrated a statistically significant reduction in missed diagnoses for high-risk conditions. In clinics using the AI engine, severe malaria cases were flagged for urgent referral an average of 2.3 days earlier than in control clinics using standard triage protocols. For sepsis, early flagging improved by 1.8 days.

These findings suggest that AI-assisted diagnostics, when appropriately calibrated and integrated into existing clinical workflows, can meaningfully improve care quality in resource-limited settings — without replacing the clinician. The system's value lies not in substituting human judgment, but in augmenting it: ensuring that no high-risk patient slips through the gaps of an overstretched system.

Limitations of this study include the reliance on clinician diagnosis as ground truth in settings where confirmatory testing is not always available, and the potential for selection bias in clinics that volunteered for AI deployment. Larger randomized controlled studies are warranted and are currently in planning.