Detecting ECG abnormalities using an ensemble framework enhanced by Bayesian belief network

Abnormality detection of Electrocardiogram (ECG) is a typical multi-label classification problem, which is often tackled by training a binary classifier for every abnormality. Unfortunately, one type of classifier usually applies to some abnormalities (labels), but cannot work well for other abnormalities, due to the classifier's pertinence and the complex feature-label correlations. Moreover, the number of abnormalities varies much for different patients (instances), which further complicates the identification of all the abnormalities. We observed that different ECG labels depend on each other to varying degrees, which is valuable for diagnosis but difficult to be captured explicitly. Thus, we present a Universal ECG Classification (UEC) framework to detect ECG abnormalities, which first constructs a more discriminating target feature space by combining the inter-labelset as well as intra-labelset features and then trains a set of binary classifiers for label voting. In particular, we propose to take advantage of a Bayesian belief network to enhance the voting of the trained binary classifiers, thus precisely determining one instance's target labelset. Experimental results on two real datasets demonstrate that our framework can effectively detect ECG abnormalities.