Delineation of the electrocardiogram with a mixed-quality-annotations dataset using convolutional neural networks

Abstract

Detection and delineation are key steps for retrieving and structuring information of the electrocardiogram (ECG), being thus crucial for numerous tasks in clinical practice. Digital signal processing (DSP) algorithms are often considered state-of-the-art for this purpose but require laborious rule readaptation for adapting to unseen morphologies. This work explores the adaptation of the the U-Net, a deep learning (DL) network employed for image segmentation, to electrocardiographic data. The model was trained using PhysioNet’s QT database, a small dataset of 105 2-lead ambulatory recordings, while being independently tested for many architectural variations, comprising changes in the model’s capacity (depth, width) and inference strategy (single- and multi-lead) in a fivefold cross-validation manner. This work features several regularization techniques to alleviate data scarcity, such as semi-supervised pre-training with low-quality data labels, performing ECG-based data augmentation and applying in-built model regularizers. The best performing configuration reached precisions of 90.12%, 99.14% and 98.25% and recalls of 98.73%, 99.94% and 99.88% for the P, QRS and T waves, respectively, on par with DSP-based approaches. Despite being a data-hungry technique trained on a small dataset, a U-Net based approach demonstrates to be a viable alternative for this task.

Generalizing electrocardiogram delineation -- Training convolutional neural networks with synthetic data augmentation

“…The ordering of operations after the convolutional operations was defined to agree with the image segmentation literature (non-linearity → batch normalization → dropout) [38,39]. All networks were trained using ECG-centered data augmentation, as described elsewhere [10], comprising additive white Gaussian noise, random periodic spikes, amplifier saturation, powerline noise, baseline wander and pacemaker spikes to enhance the model's generalizability. All executions were performed at the Universitat Pompeu Fabra's high performance computing environment, assigning the jobs to either an NVIDIA 1080Ti or NVIDIA Titan Xp GPU, and used the PyTorch library [40].…”

“…The data and ground truth, either real or synthesized, were then represented as binary masks for their usage in DL-based segmentation architectures, where a mask of shape {0, 1} 3×N was T rue-valued whenever a specific sample n ∈ N was contained within a P, QRS or T wave (indices 0, 1 and 2, respectively) and F alse-valued otherwise [10], bridging the gap with the imaging literature. Finally, the joint training database was split into 5-fold cross-validation with strict subject-wise splitting, not sharing beats or leads of the same patient in the training and validation sets [10,31]. Given that the proposed method employs pseudo-synthetic data generation, the pseudo-ECGs were also generated using data uniquely from the training set for each fold, ensuring no cross-fold contamination.…”

“…Many authors have experimented with the hyperparameters governing the U-Net, in the shape of number of convolutional operations (width), the number of upsampling-downsampling pairs (depth), starting number of convolutional filters, type of convolutional operation, type of non-linearity and presence/absence of other post-convolutional operations (batch normalization [32], spatial dropout [33]), among others, which was partially covered in [10] for the QT database.…”

“…Many computational approaches exist for the automatic quantification of the ECG. Most of these produce delineation of the electrocardiogram [4][5][6][7][8][9][10][11][12][13]. Delineation methods can be divided in two main groups: digital signal processing (DSP) and machine learning (ML) based methods.…”

“…To the best of our knowledge, these improvements have not been explored in the literature for ECG analysis. A more rudimentary version of this work exists in the literature [10]; however, the current approach displays key components that allow the algorithm to generalize better against a wider array of morphologies, such as the application of the synthetic data augmentation, the novel loss functions and a much wider array of architectural variation exploration.…”