Sürdürülebilir Kalkınma Amaçları

• 1

• MakaleAçık Erişim2026/r/gumushane/13725
  A leakage-free analytics framework for subject-wise electrocardiogram-based sleep apnea detection using machine learning

  Melek, Negin

  Sleep apnea is associated with impaired sleep quality and elevated cardiovascular risk, creating a need for reliable automated screening approaches. Although electrocardiogram (ECG)-based detection methods have shown promising results, reported performance values can be affected by evaluation strategies that allow information leakage between training and testing data. This study investigates ECG-based sleep apnea detection under strict subject-wise conditions using the PhysioNet Apnea-ECG database. Thirty-five annotated single-lead ECG recordings sampled at 100 Hz were divided into non-overlapping 60-s segments according to minute-wise apnea labels. Statistical ECG descriptors and time-domain heart rate variability features were extracted and evaluated using k-nearest neighbors (KNN), support vector machine (SVM), and Random Forest (RF) classifiers. In addition, a lightweight one-dimensional convolutional neural network (1D-CNN) was trained directly on filtered ECG segments. Repeated holdout validation with subject-level train-test separation was adopted, and normalization parameters were estimated exclusively from training data in each repetition to prevent information leakage. Model performance was assessed using accuracy, balanced accuracy, precision, recall, specificity, F1-score, ROC-AUC, Cohen's kappa, and Golden Distance (GD). In the representative subject-wise evaluation, Random Forest achieved the most balanced performance, yielding the highest balanced accuracy (0.70) together with the lowest GD (0.56). Across the repeated subject-wise holdout experiments, SVM achieved the strongest average overall performance, producing the highest mean accuracy (0.72), balanced accuracy (0.69), ROC-AUC (0.77), and Cohen's kappa (0.39), and the lowest mean GD value (0.58). The lightweight 1D-CNN converged stably during training, but its performance on unseen subjects remained lower than that of the handcrafted feature-based models.

  • 
  İncele