Addressing missing minority data in class-imbalanced medical datasets: a study of imputation and over-sampling strategies

Real-world medical datasets often suffer from both class imbalance and missing values, complicating reliable predictive modeling. While prior studies typically perform missing value imputation before over-sampling, the impact of reversing this order remains underexplored. This study compares three strategies: OS_IM (over-sampling → imputation), IM_OS_General (imputation → over-sampling using complete and imputed data), and IM_OS_Complete (imputation → over-sampling using only complete data). Experiments on 21 binary imbalanced medical datasets simulate 10%–90% missing rates in the minority class. Six imputation methods, SMOTE over-sampling, and SVM/RF classifiers are evaluated. Results show OS_IM consistently achieves lower RMSE across missing rates, indicating better imputation quality when over-sampling is applied first. For classification, IM_OS_General generally yields the best AUC at lower missing rates with RF, while IM_OS_Complete or OS_IM perform better at higher missing rates. These findings demonstrate that the order of imputation and over-sampling should be chosen based on missing rate and modeling objectives.