Classification of cardiac ECG signals using quadratic discriminant analysis for atrial fibrillation

Accurate and timely detection of atrial fibrillation (AF) is a crucial issue in avoiding the risk of serious cardiovascular complications such as stroke and heart failure. This study proposed a machine learning-based system used for automated electrocardiogram (ECG) signal classification using features derived from spectral and temporal analyses. The study used a dataset obtained from the PhysioNet 2017 challenge database with 5000 samples labeled as normal rhythm (N), AF, and noisy rhythm (NR). Five primary features: average spectral energy, variation in spectral power, root mean square fluctuation, overall signal energy, and power variance were extracted, expressing both amplitude and frequency-domain characteristics of cardiac dynamics. 10-fold stratified cross-validation approach was used to evaluate the performance of the model using multiple classifiers, which are: quadratic discriminant analysis (QDA), support vector machine (SVM), random forest (RF), K-nearest neighbors (KNN), and multi-layer perceptron (MLP). Additionally, to improve model stability and accuracy, a soft voting-based ensemble model was developed. Although RF achieved the highest accuracy (94.15%), the ensemble model (93.21%) outperformed QDA and SVM. The ensemble approach achieved sensitivities ranging from 0.89 to 0.96 and specificities between 0.95 and 0.97 across all rhythm categories, demonstrating strong generalization.