The surface electrocardiogram (ECG) is the most widely adopted test to diagnose cardiac conditions. Ambulatory ECG recordings are usually conducted throughout long spans of time of 24/48 hours, resulting in millions of data points. Automatically, locating and delineating the waves in these signals would allow clinicians to effortless extract critical biomarkers, such as heart-rate and QRS widths, from ECG data.
Although automatic QRS single lead delineation systems have been developed, the state-of-the-art is expert systems. Expert systems lack acceptance due to the complexity of adapting them to new scenarios. On the other hand, deep learning models allow the end-user to retrain them using data of their specific settings. This characteristic enables the algorithm to relearn a new optimal representation.
This study is, to the best of our knowledge, the first work to present a deep learning-based multilead ECG delineation method. This paper presents a novel 2-stepped multilead QRS delineation system based on deep learning and data augmentation techniques. The system, firstly, segments the QRS waves from the ECG recording and, then, delineates these waves individually. The system is formed by a segmentation and delineation modules, each of which is composed of two machine learning modules, namely a one-dimensional convolutional neural network and a fully connected neural network. The validation of our approach was conducted using nested cross-validation on the QT database from Physionet.
This method demonstrated to be able to successfully delineate QRS complexes when evaluated on data from different databases, in which it reached root-mean-square error performances of 3.02±0.12 and 4.63±0.27 instances for QRS onset and offset, respectively. These results were comparable to the state-of-the-art. Moreover, the presented data augmentation strategies permitted our system to successfully learn from a scarce and highly diverse set of multilead ECG recordings how to delineate QRS complexes.