Signal quality indices and data fusion for determining clinical acceptability of electrocardiograms

A completely automated algorithm to detect poor-quality electrocardiograms (ECGs) is described. The algorithm is based on both novel and previously published signal quality metrics, originally designed for intensive care monitoring. The algorithms have been adapted for use on short (5-10 s) single- and multi-lead ECGs. The metrics quantify spectral energy distribution, higher order moments and inter-channel and inter-algorithm agreement. Seven metrics were calculated for each channel (84 features in all) and presented to either a multi-layer perceptron artificial neural network or a support vector machine (SVM) for training on a multiple-annotator labelled and adjudicated training dataset. A single-lead version of the algorithm was also developed in a similar manner. Data were drawn from the PhysioNet Challenge 2011 dataset where binary labels were available, on 1500 12-lead ECGs indicating whether the entire recording was acceptable or unacceptable for clinical interpretation. We re-annotated all the leads in both the training set (1000 labelled ECGs) and test dataset (500 12-lead ECGs where labels were not publicly available) using two independent annotators, and a third for adjudication of differences. We found that low-quality data accounted for only 16% of the ECG leads. To balance the classes (between high and low quality), we created extra noisy data samples by adding noise from PhysioNet's noise stress test database to some of the clean 12-lead ECGs. No data were shared between training and test sets. A classification accuracy of 98% on the training data and 97% on the test data were achieved. Upon inspection, incorrectly classified data were found to be borderline cases which could be classified either way. If these cases were more consistently labelled, we expect our approach to achieve an accuracy closer to 100%.