A robust AF diagnosis technique using deep curve estimation and time frequency analysis on ECG

Atrial fibrillation (AF) is the most prevalent sustained cardiac arrhythmia, affecting over 33 million individuals worldwide, with incidence increasing due to aging populations and rising cardiovascular risk factors. AF is characterized by irregular atrial electrical activity that disrupts normal cardiac rhythm and significantly elevates the risk of stroke and heart failure. Early and accurate detection is critical for effective clinical intervention.Traditional machine learning (ML) techniques for AF analysis depend on handcrafted Electrocardiogram (ECG) features and often struggle to represent the non-stationary and transient nature of atrial activity. Time–frequency analysis (TFA) overcomes these limitations by jointly capturing temporal and spectral variations in 1D ECG signals, enabling enhanced characterization of AF-related patterns. In this paper, time–frequency representations, such as Wigner Ville distribution (WVD), smooth Pseudo Wigner Ville distribution (SPWVD), short-time fourier transform (STFT), Stockwell transform (ST), and Chirplet transform (CT), in combination with deep learning (DL) architectures such as spatio-temporal convolution network (STC9), dense network (DenseNet), and Deep curve estimation network (DCENet), facilitate deep feature reuse, improved gradient flow, and efficient multi-scale representation learning. These models demonstrate strong capability in automatically extracting discriminative features while reducing reliance on expert-driven preprocessing. The integration of TFA with state-of-the-art DL frameworks offers a robust pathway toward scalable and reliable AF detection for clinical and wearable monitoring applications. Moreover, the DL framework hyper-parameters are tuned. The proposed method has been validated on Physionet MIT-BIH AF datasets for analysis and it is found that the proposed method helps diagnosing AF more accurately and effectively. Among various DL-CNN techniques the customized DL networks DCENet and DenseNet with 10 epoch and 0.001 learning rate reported the highest train-test accuracy of 97.95%, 97.91%, 97.39%, 96.87%, and 95.13% for WVD, CT, ST, STFT, and SPWVD transformer. The DCENet reported the highest accuracy of 97.95% and 97.91% for WVD and CT transform, whereas DenseNet reported the highest accuracy of 97.39% and 97.22% for the ST and CT transform. The Poincare plot, nonlinear representation reported least accuracy of 76.50% for the proposed DCENet. Bootstrap confidence interval computed for WVD test images in DCENet reported mean accuracy and 95% CI are 97.34% and [96.18%, 98.38%]. Explainability of the model is analyzed by Grad-CAM and SHAP. The importance of individual block is analyzed by ablation study, whereas robustness of the model is studied for the noisy data.