Applying oversampling before cross-validation will lead to high bias in radiomics

Abstract Class imbalance is often unavoidable for radiomic data collected from clinical routine. It can create problems during classifier training since the majority class could dominate the minority class. Consequently, resampling methods like oversampling or undersampling are applied to the data to class-balance the data. However, the resampling must not be applied upfront to all data because it would lead to data leakage and, therefore, to erroneous results. This study aims to measure the extent of this bias. Five-fold cross-validation with 30 repeats was performed using a set of 15 radiomic datasets to train predictive models. The training involved two scenarios: first, the models were trained correctly by applying the resampling methods during the cross-validation. Second, the models were trained incorrectly by performing the resampling on all the data before cross-validation. The bias was defined empirically as the difference between the best-performing models in both scenarios in terms of area under the receiver operating characteristic curve (AUC), sensitivity, specificity, balanced accuracy, and the Brier score. In addition, a simulation study was performed on a randomly generated dataset for verification. The results demonstrated that incorrectly applying the oversampling methods to all data resulted in a large positive bias (up to 0.34 in AUC, 0.33 in sensitivity, 0.31 in specificity, and 0.37 in balanced accuracy). The bias depended on the data balance, and approximately an increase of 0.10 in the AUC was observed for each increase in imbalance. The models also showed a bias in calibration measured using the Brier score, which differed by up to −0.18 between the correctly and incorrectly trained models. The undersampling methods were not affected significantly by bias. These results emphasize that any resampling method should be applied correctly only to the training data to avoid data leakage and, subsequently, biased model performance and calibration.