Lightweight CNN for multi-class ECG arrhythmia detection using spectrogram-based time–frequency representations

Objective This study proposes a lightweight yet accurate framework for electrocardiogram (ECG) arrhythmia classification that is suitable for real-time deployment in wearable devices, embedded systems, and edge-AI healthcare applications. Methods A lightweight two-dimensional convolutional neural network (2D CNN) was developed using high-resolution short-time Fourier transform (STFT) spectrograms generated from ECG segments containing five consecutive R-peaks. This physiologically meaningful segmentation strategy preserves both heartbeat morphology and local rhythm information. The proposed CNN consists of six convolutional blocks with progressive feature extraction and max-pooling operations. Performance was evaluated under both the conventional train–validation–test protocol and a record-wise three-fold cross-validation protocol. Comparative analyses were conducted against representative deep CNN architectures and lightweight models, including VGG16, VGG19, ResNet variants, Inception-ResNet V2, DenseNet, Autoencoder-based networks, MobileNetV2, EfficientNet-B0, and ShuffleNetV2. Results Under the conventional train–validation–test protocol, the proposed CNN achieved an accuracy of 86.16%, an F1-score of 86.07%, and a specificity of 96.54% while requiring only 898,413 trainable parameters. Under the record-wise three-fold cross-validation protocol, the proposed model achieved the highest macro AUROC among the evaluated lightweight CNNs while requiring only 0.71 GFLOPs and demonstrating low inference latency. Ablation experiments showed that the high-resolution STFT representation substantially improved classification performance over raw one-dimensional ECG input, while five-consecutive-R-peak segmentation outperformed single-beat segmentation and preserved physiologically meaningful temporal context. Conclusions The proposed framework provides an effective balance between classification performance and computational efficiency by integrating physiologically meaningful heartbeat segmentation, high-resolution time–frequency representation, and a lightweight CNN architecture. These characteristics make it a promising solution for real-time ECG arrhythmia classification in wearable cardiac monitoring devices, embedded intelligent healthcare systems, and edge-AI applications.