The Temporal Fusion Transformer within the CEDRC-Net framework achieved 98% accuracy and lower reconstruction errors in ECG signal reconstruction compared to the Variational Autoencoder.
Does the CEDRC-Net using Temporal Fusion Transformer improve ECG signal reconstruction and downstream cardiovascular disease classification compared to Variational Autoencoder?
The CEDRC-Net framework using Temporal Fusion Transformer significantly improves ECG denoising, missing data reconstruction, and downstream classification accuracy for cardiovascular diseases.
Absolute Event Rate: 98% vs 96%
Abstract Cardiovascular diseases (CVDs) are a significant and widespread cause of death in the world, continuing to increase mortality rates. Therefore, timely identification and diagnosis are essential for a patient’s optimized recovery and longevity. In this regard, ECG is an effective tool for detecting anomalous heart conditions. However, interfering factors like noise, transient changes, and missing data could affect the accurate diagnosis of CVDs. The Cardiovascular ECG Data Recovery and Classification Network (CEDRC-Net) sorts and collates noisy data while also recovering missing data caused by equipment malfunction or human error, using a multistage machine-learning and deep-learning model. Moreover, CEDRC-Net is incorporated into the Transformer-based Convolutional Denoising Autoencoder (TCDAE) model to methodically mitigate noise, subsequently the Variational Autoencoder (VAE) or Temporal Fusion Transformer (TFT) are systematically employed as an alternative to accurately reconstruct and forecast ECG signals. Following this, the system classified heart diseases, including atrial fibrillation, sinus bradycardia, and tachycardia, using several machine-learning algorithms, based on data from a dataset comprising 2426 patients.TFT showcased better performance than VAE in ECG signal reconstruction, achieving up to 98% accuracy with Gradient Boosting, compared to 96% by VAE. Furthermore, in downstream classification, signals enhanced by TFT led to superior model results, with SVM and XGBoost both reaching 98.4% accuracy and F1-scores. The TFT achieved substantially lower reconstruction errors (MAE = 0.015, MSE = 0.00045, RMSE = 0.0132) compared to the VAE (MAE = 0.075, MSE = 0.011, RMSE = 0.107). These results highlight TFT’s strong denoising capability for improving ECG diagnostic accuracy, while the suggested system ensures reliable measurements under noisy and non-ideal conditions. It is highly conducive for early and accurate diagnosis, better clinical decisions, and decreased patient load on the cardiologists. The research is performed on the MIMIC-IV-ECG 12-lead real-time dataset.
Khan et al. (2026) studied Cardiovascular diseases (atrial fibrillation, sinus bradycardia, tachycardia) (n=2,426). Temporal Fusion Transformer (TFT) within CEDRC-Net vs. Variational Autoencoder (VAE) was evaluated on ECG signal reconstruction accuracy. The Temporal Fusion Transformer within the CEDRC-Net framework achieved 98% accuracy and lower reconstruction errors in ECG signal reconstruction compared to the Variational Autoencoder.