CLASS-IMBALANCE METHODS IN REGRESSION-BASED ANAEMIA SCREENING: A COMPARATIVE EVALUATION

Abstract

Imbalance corrections are known to harm calibration without improving discrimination in logistic risk models, but this has not been examined where a regressor predicts a continuous marker and categories follow from diagnostic cut-offs. Using 103 pregnant women, we compared class weighting, focal-style weighting, SMOTER and binary collapse against an unmodified baseline and a decision-threshold shift. Across 50 matched folds no method changed AUC significantly, while threshold shifting produced the largest sensitivity gain.