A deep learning framework fusing spectrograms and scalograms with convolutional neural networks, bidirectional long short-term memory, and an attention mechanism achieved 93.4 percent accuracy in classifying record-derived electrocardiogram segments, according to research published in the cardiac electrophysiology section of Frontiers in Physiology. The supervised three-class investigation examined whether distinct morphological and transformed-domain characteristics could differentiate cardiac signal windows extracted from electrocardiograms labelled as normal rhythm, atrial fibrillation, or premature ventricular contractions.
To perform the supervised classification task, the research team analyzed recordings obtained from the open PTB-XL database. The data processing workflow incorporated digital filtering, signal normalization, and automated R-peak detection carried out using the Pan–Tompkins algorithm. Following these preparatory steps, the authors extracted two-second, twelve-lead electrocardiogram segments with dimensions of 1,000 time points across 12 channels. These multi-lead windows were centered directly on detected R-peaks, creating consistent matrix representations of the cardiac signals for subsequent algorithmic processing.
Because the extracted two-second segments were centered on detected R-peaks, the investigators noted that each window could contain multiple cardiac cycles rather than representing a single isolated heartbeat. Crucially, every extracted segment inherited the diagnostic label assigned to its overarching source recording in the database. As a result, segments designated as atrial fibrillation-associated reflected the general classification of the source record. The authors cautioned that this labeling scheme means the model’s classifications must not be interpreted as independent diagnostic determinations of atrial fibrillation from isolated beats.
To evaluate time-frequency characteristics alongside raw morphology, the researchers converted the extracted twelve-lead electrocardiogram segments into two distinct visual domain representations by generating both spectrograms and scalograms. These complementary transformed inputs were then classified using a hybrid neural network architecture.
When evaluated on an independent held-out test set, the proposed CNN–BiLSTM–attention framework demonstrated superior classification accuracy relative to simpler model variants evaluated under identical conditions. The full attention-augmented architecture attained a top accuracy of 93.4 percent on the independent test data. By comparison, an ablated CNN–BiLSTM model without the attention mechanism achieved an accuracy of 91.8 percent, while baseline architectures consisting of a convolutional neural network alone and a bidirectional long short-term memory network alone achieved lower accuracies of 88.6 percent and 86.9 percent, respectively.
The researchers examined the optimization behavior of the neural network over 300 training epochs, noting that complete convergence could not be definitively established from the recorded validation trajectories. The study was conducted by Kavimbi Chipusu from the University of Saskatchewan in Canada alongside Guangfeng Li, Zhidong Zhang, Gang Qiao, Xiaosan Chen, and Gangqiang Zhou from Central China Fuwai Hospital of Zhengzhou University in China. The team concluded that while spectrogram–scalogram fusion effectively classifies record-derived normal, atrial fibrillation, and premature ventricular contraction segments, clinical application requires recognizing that record-inherited labels do not constitute beat-level diagnoses.
Subject of Research: Electrocardiogram segment classification using deep learning and time-frequency transforms
Article Title: Spectrogram–scalogram fusion with CNN–BiLSTM– attention for record-derived ECG segment classification
Article References: Zeng, T., Zeng, Y., Li, G., Zhang, Z., Chipusu, K., & Huang, J. (2026). Spectrogram–scalogram fusion with CNN–BiLSTM– attention for record-derived ECG segment classification. Frontiers in Physiology, 17, Article 1854203. https://doi.org/10.3389/fphys.2026.1854203
Image Credits: AI Generated
DOI: 10.3389/fphys.2026.1854203
Keywords: electrocardiogram, arrhythmia classification, spectrogram, scalogram, CNN-BiLSTM-attention, PTB-XL dataset, atrial fibrillation, premature ventricular contraction, Pan-Tompkins algorithm, R-peak detection, 12-lead ECG, cardiac electrophysiology, deep learning, bidirectional LSTM, convolutional neural networks, attention mechanism
News Source: Blake Davidson. (November 19, 2026). Deep Learning Fusion Architecture Accurately Classifies Record-Derived ECG Segments. Scienmag.



