Deep learning from 1 million clinical ecgs using transform domain representations for arrhythmia classification and ejection fraction estimation

Background Clinical ECGs contain complex temporal and spectral patterns that extend beyond human visual interpretation. Deep learning can capture these multidimensional features to detect both rhythm disturbances and ventricular dysfunction. Combining complementary time- and frequency-based representations may enhance diagnostic accuracy and generalizability across ECG sources. Purpose To develop and validate deep learning models trained on clinical ECGs for (1) multiclass arrhythmia classification and (2) estimation of reduced left ventricular ejection fraction (LVEF < 40%), using real-time, Fourier, and Stockwell transform representations. Methods A total of 1 000 000 clinical ECGs from a large healthcare system were analyzed. Of these, 725 000 ECGs were used for arrhythmia classification across 18 rhythm and conduction categories, and 275 000 ECG–LVEF pairs for ejection fraction estimation. Each ECG was represented in three complementary domains: (1) time domain (TD) preserving waveform morphology and conduction timing, (2) Fourier domain (FD) capturing global spectral features, and (3) Stockwell domain (TFD) encoding both temporal and spectral localization. Separate convolutional deep learning models were trained for each representation using four convolutional and two dense layers (ReLU activation, Adam optimizer, learning rate = 0.0001). Internal validation used held-out clinical ECGs; external validation employed open-access databases and additional ECGs in PDF format from independent institutions. Results For arrhythmia classification, mean AUC/F1-scores on internal validation were 0.933/0.883 (TD), 0.949/0.887 (FD), and 0.963/0.903 (TFD). External validation achieved 0.894/0.808, 0.920/0.828, and 0.933/0.852, respectively. For LVEF classification, AUC ranged 0.85–0.92, with the TD model achieving AUC = 0.858 (accuracy 0.776). Validation on an independent clinical cohort of atrial fibrillation ECGs achieved AUC = 0.997, confirming strong generalizability. Across both tasks, the Stockwell (time–frequency) model consistently performed best, highlighting the benefit of joint temporal–spectral learning. Conclusions Deep learning models trained on 1 million clinical ECGs can accurately classify arrhythmias and detect reduced LVEF. Incorporating time-, frequency-, and time–frequency-domain representations improve feature extraction and model robustness across datasets, supporting the role of domain-diverse deep learning in next-generation AI-enabled ECG diagnostics.