Reduced features set neural network approach based on high-resolution time-frequency images for cardiac abnormality detection.

A suitable temporal and spectral processing of the electrocardiogram (ECG) signals can facilitate the visual interpretation and discrimination between known patterns for classification. This paper proposes a non-invasive hybrid neural network and time-frequency (TF) based method to detect and classify commonly found cardiac abnormalities in ECG signals including congestive heart failure, ventricular tachyarrhythmia, intracardiac atrial fibrillation, arrhythmia, malignant ventricular ectopy, normal sinus rhythm, and postictal heart rate oscillations in partial epilepsy. Non-stationary raw ECG signals are collected from an online healthcare dataset source 'PhysioBank' that contains physiologic signals. These temporal signals are processed through Wigner-Ville distribution to produce high-resolution and concentrated TF images depicting specific visual patterns of cardiac abnormalities. The TF images are used to extract the abnormality parameters with the help of medical experts with good diagnostic accuracy. Principal component analysis (PCA) is employed for feature reduction and important features selection from the ECG signals. The selected features are used for training the multilayer feed-forward artificial neural network (ANN) for detection and classification while training parameters like the number of epochs, activation functions, and the learning rate is suitably selected with appropriate stopping criteria. Experimental results demonstrate the effectiveness of the hybrid neural-TF approach using PCA for abnormality detection and classification.