Cardiac Signal Abnomaly Detection Based On Deep Convolutional Neural Network

Cardiac auscultation and electrocardiogram(ECG)is the most practiced non-invasive and cost-effective procedure for the early diagnosis of heart diseases.However,the severe shortage of trained physicians and health workers in some underserved communities who can perform cardiac diagnosis exacerbates the threat as early diagnosis of Cardiocascular diseases become unlikely.This demands the development of machine learning-based assistive technologies for cardiac screening.Due to the influence of environment,the divergence of different digital stethoscopes,and data collection protocol,the pattern of heart sound signals are so complex that fixed pattern feature extraction or learning features directly from the signal can not enough lead to final accurate classification.For this issue,a learnable lifting wavelet transform block(Le-LWT),which embeds the trainable convolutional neural network into the lifting wavelet transform,is proposed in this thesis.Le-LWT can utilize the non-linear learning ability of CNN while maintaining the multi-resolution time-frequency analysis ability of wavelet transform,as well as more interpretation than the deep networks designed as black boxes.Based on Le-LWT module,we propound a lightweight end-to-end Le-LWTNet that has stronger non-linear characterization capabilities and few parameters for automatic abnormality detection of the heart sound.Experimental evaluations are performed on a 10-fold cross-validation task using the2016 Physio Net/Cin C Challenge dataset and the new publicly available pediatric heart sound dataset we collected.Results demonstrate that the proposed method excels the state-of-the-art models both in abnormality detection and parameter consumption.In particular,the proposed network excels in identifying abnormal heart sounds with F1-score above 90% and a specificity over 98%.Computerized ECG interpretation plays a critical role in the clinical ECG workflow.However,there is limited evidence to demonstrate the utility of reduced-lead ECGs for capturing a wide range of diagnostic information.Thus,in this work,we aimed to develop a deep learning-based algorithm to identify clinical diagnoses from twelve-lead,sixlead,four-lead,three-lead,and two-lead ECG recordings.We have employed a branched DNN that utilizes two CNNs(similar to the residual network,but adapted to signals with varying dimensions)with different filter sizes.This architecture can effectively capture abnormal patterns of diseases and suppress noise interference by introducing double branch and skip connections.Moreover,the Squeeze-and-Excitation block is added to every branch of the architecture,which allows the DNNs to learn more effective features from input signals.In this work,we demonstrate our branched network on an open dataset,the experimental results convince the effectiveness of the proposed method.