Systemic modelling for atrial fibrillation detection: integrating MobileNetV2 transfer learning with Bayesian-optimised KNN

Atrial fibrillation (AFB) is a leading cause of life-threatening heart diseases, and its increasing prevalence in recent times has sparked interest in the development of accurate and reliable detection algorithms. An electrocardiogram (ECG) is the most trusted tool to detect these cardiac disorders. However, computer-aided algorithms are becoming increasingly important for efficient and timely detection. This work presents a novel algorithm to detect AFB by employing a transfer-learning approach and a Bayesian optimised K-nearest neighbour (KNN) classifier after segmentation of the ECG signal into four-second segments. Moreover, the signal-to-image conversion uses continuous wavelet transform and Stockwell transform, followed by modified pre-trained deep convolutional neural network (CNN) models to extract potential attributes. The extracted attributes are fed further to a feature selection technique that utilises fuzzy entropy to assess their relevance and, finally, sent to a Bayesian optimised KNN classifier to classify into normal, atrial fibrillation, and atrial flutter classes.