Enhancing cardiotocography classification via ensemble learning and threshold optimization

Machine learning classifiers trained on imbalanced healthcare datasets often exhibit bias, leading to poor performance on critical cases. The cardiotocography (CTG) dataset exemplifies this issue, where misclassification of pathological cases arises due to both class imbalance and non-optimal probability thresholds. Statistical analysis suggests refining classification thresholds, but this approach has been largely overlooked in CTG data research. To address these challenges, we propose a multifusion method integrating undersampling, threshold-moving optimization, and ensemble classifiers to enhance classification precision while maintaining computational efficiency. Applied to a CTG dataset of 502 cases from Czech Technical University and University Hospital Brno, our method showed significant improvements in identifying pathological cases. While baseline models correctly classified only about 2 out of 11 cases per test, our approach achieved 76.92, 75, and 41.67% precision, accurately identifying 9, 9, and 3 cases out of 12, respectively.