Arrhythmias detection using ECG with deep learning and superlet transform

Abstract Accurate arrhythmia detection is clinically crucial, yet performance often degrades across subjects, devices, and noise conditions. Recent studies using short-time Fourier transform and continuous wavelet transform are frequently encountered with electrocardiogram signals, limiting the ability to detect complex arrhythmias like Congestive Heart Failure (CHF), Atrial Fibrillation (AF), and Ventricular Ectopy (VE). This study presents the contribution of time–frequency representations for arrhythmia classification and provides a reproducible, subject-wise evaluation that attenuates class imbalance. Electrocardiogram signals were acquired from PhysioNet and pre-processed with a fourth-order Butterworth bandpass and segmented into 5-second windows. One-dimensional signals into two-dimensional time-frequency representations were produced by the Superlet transform (SLT). The SLT-transformed images were fed to three Convolutional Neural Network (CNN) architectures—MobileNet, InceptionV3, and VGG16, and were trained using five-fold cross-validation for 30 epochs. These images were resized to 227 × 227 × 3 pixels, and data augmentation was performed to mitigate class imbalance, resulting in 39,123 images. The VGG16 model attained the highest accuracy at 98.73%, followed by InceptionV3 at 97.55% and MobileNet at 98.26%. Strong robustness and generalizability were demonstrated by these models across different arrhythmia categories. The combination of the SLT and deep neural networks makes it an extremely powerful tool for arrhythmia detection. These findings underscore the integration between advanced time-frequency transformations and sophisticated deep learning strategies in identifying complex cardiac anomalies, which can expedite clinical workflows and improve patient outcomes. Further research into the interpretability of these models may offer deeper clinical insights.