Cardiotocographic Signal Classification: Ensemble Classifier Assisted by Probability Threshold Optimization on Imbalanced Ctg-Hub Data

Abstract

In many real-world scenarios, the focus is often on the minority class, where machine learning classifiers trained on imbalanced datasets tend to exhibit bias, resulting in poor performance on the minority class. The Cardiotocography (CTG) dataset is one such example. Previous research has shown that imbalanced data can hinder model training, leading to misclassifying pathological cases. Moreover, the default probability threshold is set to 0.5, which may not be optimal for such medical datasets.A statistical analysis suggests selecting a threshold beyond which new observations are classified as 1 or 0, transitioning from statistics to decision-making. However, this issue has received limited attention in this type of data. To address this gap, we propose a multi-fusion approach that includes undersampling to mitigate data imbalance, threshold moving optimization for improved classification, and ensemble classifiers for enhanced performance and noise resilience. This study applies the method to the CTG dataset from the Czech Technical University in Prague and the University Hospital in Brno.We conducted experimental evaluations on 502 valid CTG data points and compared the outcomes with baseline classifiers. The results illustrate the effectiveness of our approach based on Stratified K-Fold cross-validation, resulting in improved classification accuracy for pathological cases. Specifically, the baseline models correctly classified around 2 cases out of 11 in each test, whereas the proposed model achieved around 76.92\%, 75 \%, and 41.67\% accuracy in three tests with correctly classified 9, 9, and 3 cases increase out of 12 than the baseline classifier on pathological cases, respectively.

Keywords: Cardiotocograph, Hypoxemia, Ensemble classifier Moving threshold Probalistic random forest UnderSampling dataset