Heart disease diagnosis with ECG signal image and augmented feature set using improved LeNet model

Heart disease is a leading cause of mortality all over the world. Electrocardiogram (ECG) is the less expensive non-invasive procedure among the diverse diagnosis techniques. However, it also faces challenges like the complexity of ECG interpretations, scarcity of medical experts, heart disease comorbidity and heart disease’s manifestation similarities in ECG signals. Machine learning and Deep Learning algorithms are feasible alternatives to the traditional heart disease diagnosis methods. For that reason, we have proposed hybrid Improved LeNet with LinkNet model to detect the heart disease using ECG signal images, which includes 5 phases like pre-processing, feature extraction, feature selection, data augmentation and classification. Initially, the input ECG signal images were pre-processed or denoised using bilateral filtering technique. Subsequently, features like colour, hierarchy of skeleton, Improved Local Gradient Patterns, deep features and statistical features were extracted from the denoised images. Then, from these retrieved features, most relevant or suitable features were chosen using Hybrid Improved Filter and Wrapper-based feature selection technique. Then these chosen features subjected to augmentation using random sampling process. Finally, this augmented feature is subjected to classification phase, where a hybrid model (ILN-LNK), that is the combination of Improved LeNet (ILN) and LinkNet (LNK) is utilized for heart disease classification. The proposed method was evaluated on the Mendeley ECG Images Dataset of Cardiac Patients. Moreover, this ILN-LNK model obtained superior accuracy and sensitivity of 0.991 and 0.975 at 90% training data.