Key result
Normalized cosine-similarity head improves MobileViTv2 ECG classification accuracy to ~98% over standard linear head.
Why the study?
The study aimed to classify heartbeats from the MIT-BIH Arrhythmia Database into the five AAMI classes using a lightweight hybrid convolution-transformer network while mitigating class imbalance and examining model errors.
Does replacing a standard linear classification head with a normalised cosine-similarity head in a MobileViTv2 network trained on continuous wavelet transform images and synthetically balanced data improve five-class ECG heartbeat classification accuracy and macro F1-score on the MIT-BIH Arrhythmia Database?
Does replacing a standard linear classification head with a normalised cosine-similarity head in a MobileViTv2 network trained on continuous wavelet transform images and synthetically balanced data improve five-class ECG heartbeat classification accuracy and macro F1-score on the MIT-BIH Arrhythmia Database?
Replacing a standard linear classifier with a normalised cosine-similarity head in a MobileViTv2 network trained on CWT images and synthetic GAN/autoencoder-balanced beats modestly improved intra-patient ECG arrhythmia classification (accuracy 0.9793 vs 0.9745), though patient-independent validation is still required before clinical or wearable application.
Answers are grounded in this paper and cite their sources.
May modestly advance automated ECG classification; leaves open prospective clinical validation before practice adoption.
Purpose To classify heartbeats from the MIT-BIH Arrhythmia Database into the five AAMI classes with a lightweight hybrid convolution–transformer network, while mitigating class imbalance and examining the model’s errors. Methods Beats were segmented around the R peak, converted to continuous wavelet transform (CWT) images and used to train MobileViTv2 with its linear classifier replaced by a normalised cosine-similarity head (NormHead). The training set was balanced to 4,000 beats per class using beats synthesised by a Wasserstein GAN with gradient penalty (WGAN-GP) and a convolutional autoencoder; the test set (7,712 beats) contained only real beats. Grad-CAM was used for qualitative error analysis. Results On a stratified, beat-level hold-out set the model reached an accuracy of 0.9793 and a macro F1-score of 0.9538 (0.9795 and 0.9515 in a second run), against 0.9745 and 0.9418 for the same backbone with a standard head. Classes F and SVE remained the hardest (F1-scores of 0.8812 and 0.9267). Synthetic beats were highly similar to real beats in time and frequency for most classes. Conclusion The cosine-similarity head gave a modest improvement over a standard head in this setting. Because training and test beats were split at the beat level and not by patient, the results describe intra-patient performance and are not evidence of generalisation to unseen patients; patient-independent and cross-database validation are still required.
No takes yet. Share an insight, caveat, or question.
Silva et al. (2026) studied Cardiac arrhythmia (n=7,712). MobileViTv2 with normalised cosine-similarity classification head (NormHead) and synthetic data balancing vs. MobileViTv2 with standard linear classification head was evaluated on ECG beat classification accuracy on held-out test set. MobileViTv2 with a normalised cosine-similarity head achieved 97.93% accuracy and a 0.9538 macro F1-score versus 97.45% and 0.9418 for the standard head on MIT-BIH ECG beats.
Synapse has enriched 10 related papers. Consider them for comparative context: