Threshold and Calibration Effects in Medical Machine Learning

Abstract

Decision thresholding and probability calibration play a key role in medical machine learning, where predicted probabilities are interpreted as risk estimates. Prior studies have shown that strong discrimination does not guarantee well-calibrated probabilities and that threshold-based decisions depend on operating conditions. This study experimentally examines the effects of probability calibration and decision thresholding on structured medical datasets. Multiple representative classifiers are evaluated using fixed and data-driven thresholding strategies. Threshold-dependent performance is analyzed together with expected calibration error. The results show that data-driven thresholds can improve balanced accuracy. Calibration effects differ across models and datasets. Isotonic calibration markedly reduces calibration error for multilayer perceptrons, whereas effects for logistic regression and random forests are smaller and inconsistent. These findings indicate that calibration and thresholding should be examined together rather than inferred from discrimination alone.

Ähnliche Arbeiten