Machine Learning Generalizability in Cervical Cancer Diagnosis: A Study Protocol Comparing Standard Benchmark and African-Collected Datasets

This protocol describes a two-arm computational study examining whether machine learning models trained on standard, non-African cervical cancer cytology benchmark datasets (SIPaKMeD, Herlev) generalize to an independently collected African cervical imaging dataset (Malhari, hosted on Mendeley Data). A second arm uses a Zambian clinical registry dataset to identify predictors of late-stage cervical cancer diagnosis. All datasets used are open-access and require no registration or data-use agreement. This is a pre-results study protocol; a follow-up manuscript reporting experimental findings will be linked upon completion.