Empirical Investigation of Class Imbalance Correction Strategies for Healthcare Diagnosis Prediction in Nigeria

Class imbalance in medical datasets, where rare conditions represent a small proportion of records, is a pervasive challenge in healthcare predictive modelling and its handling has outsized effects on clinical utility in Nigerian health AI applications. This study empirically investigated the effectiveness of seven class imbalance correction strategies for predicting rare adverse outcomes in healthcare datasets from two South South Nigerian hospitals: University of Benin Teaching Hospital and Federal Medical Centre Yenagoa. Three clinical prediction tasks were used as benchmarks: sepsis outcome within 48 hours, severe anaemia in obstetric patients, and post-operative complication occurrence, with positive class rates of 7.3, 11.4, and 9.8 percent respectively. Seven strategies were compared: no correction, random oversampling, SMOTE, ADASYN, random undersampling, NearMiss, and a cost-sensitive learning approach. Models were gradient boosting classifiers evaluated on G-mean and clinical F2-score prioritising recall. Cost-sensitive gradient boosting achieved the best clinical utility across all three tasks, with G-mean improvement of 0.18 over the unbalanced baseline. ADASYN showed the best performance on the sepsis task specifically. SMOTE underperformed ADASYN on all three tasks. Random undersampling showed poor overall performance due to information loss. The study fills an empirical gap in imbalance correction benchmarking for Nigerian clinical datasets and recommends cost-sensitive learning as the default approach for healthcare prediction models, with sensitivity analysis of correction strategies required before deployment.

Keywords: class imbalance, healthcare prediction, SMOTE, cost-sensitive learning, Nigerian hospitals