Unverified Vendor Claims and Preventable Harms: A Mixed-Methods Longitudinal Independent Audit of Health AI System Performance in Nigeria

ABSTRACT Objective To independently audit vendor-reported performance claims of health AI systems deployed in Nigeria and assess discrepancies, clinical consequences, equity impacts, and implications for safe AI deployment in low- and middle-income countries. Methods and analysis We conducted a mixed-methods longitudinal audit (October 2024-March 2026) of six health AI systems (chest X-ray interpretation, TB screening, symptom triage, maternal health risk prediction, patient history intake, and health chatbots) across 73 diverse health facilities in six Nigerian states, involving 52,000 patients and 45 key informant interviews conducted with stakeholders. All data were sourced from integrated facility-level records, and no database linkage was performed. Vendor claims were abstracted from documentation, white papers, and validation studies. Independent performance was verified by an independent third party through system logs, patient records, clinical outcomes, and stakeholder interviews. Performance gaps were quantified as absolute percentage-point differences; clinical harms were estimated using patient volume and bootstrap confidence intervals; equity impacts were assessed across vulnerability dimensions (geography, age, income, comorbidities, infrastructure) using interaction terms in mixed-effects models and an Equity Harm Index (EHI). Results Vendor-reported accuracy averaged 91.5%, while independently measured real-world accuracy averaged 67.3%, yielding a mean performance gap of 24.2 percentage points (95% CI: 21.5 to 26.9; p<0.001) across systems. Gaps ranged from 17 to 35 percentage points and were statistically significant for all systems. These discrepancies translated to substantial preventable harm, including an estimated 1,247 undetected TB cases (186 preventable deaths) and 342 misclassified high-risk pregnancies annually. Performance gaps were 28-38% larger among vulnerable groups (e.g., rural patients showed 38% higher EHI). Gaps were classified as systematic, context-dependent, or population-dependent. Conclusion Vendor-reported performance metrics substantially overstated the real-world effectiveness of health AI in Nigeria, leading to preventable patient harm and widening inequities. Mandatory independent post-deployment verification, analogous to pharmaceutical Phase IV surveillance, is essential to ensure safe, equitable AI use in resource-constrained settings. Donors and regulators should prioritize verification over trust-based deployment. KEY MESSAGES What is already known on this topic Health AI systems are rapidly deployed in LMICs, mainly based on vendor-reported performance from controlled or retrospective validations, often accepted without independent real-world checks. Evidence suggests that discrepancies arise from domain shifts, data quality issues, and population differences, but large-scale, independent audits in LMIC settings are scarce, creating a ‘verification paradox’. What this study adds This audit of six health AI systems across 73 Nigerian facilities and 52,000 patients found a consistent 24.2 percentage-point gap (95% CI 21.5-26.9) between vendor-reported accuracy (mean 91.5%) and real-world performance (mean 67.3%), causing preventable harms including about 1,247 undetected TB cases, 186 preventable deaths, and 342 misclassified high-risk pregnancies annually. The gaps were 28-38% larger in vulnerable groups (e.g., 38% higher Equity Harm Index in rural patients) and classified as systematic, context-dependent, or population-dependent. How this study might affect research, practice or policy Vendor self-reporting is unreliable, and mandatory independent post-deployment verification, similar to pharmaceutical Phase IV surveillance, is essential for patient safety and equity in global health AI. Donors, governments, and regulators should adopt verification-based models to prevent harm, reduce inequalities, and inform future multinational studies and LMIC regulatory frameworks.