FMCEB-CXR: A Structured Chest X-ray Benchmark from Nigeria for Evaluating AI Generalization in African Clinical Settings

FMCEB-CXR is the first structured chest X-ray benchmark from a Nigerian hospital, designed for evaluating cross-domain generalization of chest X-ray AI models on African clinical data. The dataset comprises 5,517 radiology reports from Federal Medical Centre Ebute-Metta, Lagos, Nigeria, with multi-label annotations across 14 categories (13 pathologies plus Normal) extracted using a hybrid NLP pipeline combining a 447-phrase bilingual medical dictionary with fine-tuned BioClinicalBERT. Disease prevalence differs substantially from Western benchmarks: Cardiomegaly 30.6% (vs 12.3% CheXpert), Normal 42.4% (vs 8.9%), Pleural Effusion 13.8% (vs 40.3%). Version 1.0 (current): NLP-extracted labels with confidence scores and extraction code. Clinician-validated labels will be added in Version 1.1. Ethics approval: FMC Ebute-Metta Research Ethics Committee, March 2026.