DETECTION OF CONGESTIVE HEART FAILURE USING MULTI-LEAD ECG SIGNAL ANALYSIS

Abstract

Congestive heart failure (CHF) is a condition in which the heart is unable to pump blood as well as it should to meet the needs of normal circulation. Over time this condition can decrease physical capacity, increase hospital visits and increase the risk of serious cardiac events. The early diagnosis of CHF is important for patient care and follow-up monitoring because of these consequences. ECG recording is a prac tical tool for this purpose, as it is inexpensive, non-invasive and suitable for repeated measurement. Multi-lead ECG data are especially useful, because different leads ob serve the electrical behaviour of the heart from different directions. Most conventional methods are hand-crafted measures and may not well represent morphological changes of ECG with time. Furthermore, evaluation on random segment splitting may provide an optimistic estimate, if the signals from one patient are present in the training and testing set. This dissertation investigates a multi-lead diagnostic attention-based re current neural network (MLDA-RNN) for classification of CHF. We use the BIDMC Congestive Heart Failure Database and the MIT-BIH Normal Sinus Rhythm Database. The data is separated patient wise, so that testing is done on unseen subjects. Inte ger Haar Wavelet Transform is included to get compact signal information into low computational cost. The MLDA-RNN model achieved an accuracy of 96.06 % on the blind test data. The result shows that the selected combination of patient-wise eval uation, multi-lead ECG representation, recurrent sequence learning, attention-based weighting and efficient wavelet processing can provide a solid framework for auto mated CHF screening. The approach can be further extended for decision-support tools, remote monitoring workflows and wearable ECG systems after broader clinical validation.