ECG Beat Classification using Optimal Projections in Overcomplete Dictionaries

Wearable health monitoring systems (WHMS) enable ubiquitous and unobtrusive monitoring of a variety of vital signs that can be measured non-invasively. These systems have the potential to revolutionize healthcare delivery by achieving early detection of critical health changes and thus possibly even disease or hazardous event prevention. Amongst the patient populations that can greatly benefit from WHMS are Congestive Heart Failure (CHF) patients. For CHF management the detection of heart arrhythmias is of crucial importance. However, since WHMS have limited computing and storage resources, diagnostic algorithms need to be computationally inexpensive. Towards this goal, we investigate in this paper the efficiency of the Matching algorithm in deriving compact time-frequency representations of ECG data, which can then be utilized from an Artificial Neural Network (ANN) to achieve beat classification. In order to select the most appropriate decomposition structure, we examine the effect of the type of dictionary utilized (stationary wavelets, cosine packets, wavelet packets) in deriving optimal features for classfication. Our results show that by applying a greedy algorithm to determine the dictionary atoms that show the greatest correlation with the ECG morphologies, an accurate, efficient and real-time beat classification scheme can be derived. Such an algorithm can then be inexpensively run on a resource-constrained portable device such as a cell phone or even directly on a smaller microcontroller-based board. The performance of our approach is evaluated using the MIT-BIH Arrhythmia database. Provided results illustrate the accuracy of the proposed method (94.9%), which together with its simplicity (a single linear transform is required for feature extraction) justify its use for real-time classification of abnormal heartbeats on a portable heart monitoring system.