Deep Learning for ECG Analysis: Benchmarks and Insights from PTB-XL

Abstract

Electrocardiography (ECG) is a very common, non-invasive diagnostic procedure and its interpretation is increasingly supported by algorithms. The progress in the field of automatic ECG analysis has up to now been hampered by a lack of appropriate datasets for training as well as a lack of well-defined evaluation procedures to ensure comparability of different algorithms. To alleviate these issues, we put forward first benchmarking results for the recently published, freely accessible clinical 12-lead ECG dataset PTB-XL, covering a variety of tasks from different ECG statement prediction tasks to age and sex prediction. Among the investigated deep-learning-based timeseries classification algorithms, we find that convolutional neural networks, in particular resnet-and inception-based architectures, show the strongest performance across all tasks. We find consistent results on the ICBEB2018 challenge ECG dataset and discuss prospects of transfer learning using classifiers pretrained on PTB-XL. These benchmarking results are complemented by deeper insights into the classification algorithm in terms of hidden stratification, model uncertainty and an exploratory interpretability analysis, which provide connecting points for future research on the dataset. Our results emphasize the prospects of deep-learning-based algorithms in the field of ECG analysis, not only in terms of quantitative accuracy but also in terms of clinically equally important further quality metrics such as uncertainty quantification and interpretability. With this resource, we aim to establish the PTB-XL dataset as a resource for structured benchmarking of ECG analysis algorithms and encourage other researchers in the field to join these efforts.

Automated Comprehensive Interpretation of 12-lead Electrocardiograms Using Pre-trained Exponentially Dilated Causal Convolutional Neural Networks

“…Earlier studies showed that this technique outperforms recurrent neural networks, another technique that allows for increasing receptive fields, in terms of both efficiency and prediction performance [6]. Furthermore, Strodthoff et al [7] have shown that transfer learning can be applied to improve ECG classifiers. Likewise, we have demonstrated that transfer learning, using the data available at the UMCU, to a dataset recorded using a different ECG device and acquired from an ethnically and geographically different population, can be effective for improving ECG classification.…”

“…These convolutions have the advantage that they take the temporal nature of the ECG into account, while efficiently learning long-range dependencies in time series [6]. Furthermore, other works have discussed the prospects of using large task-related ECG datasets to pre-train ECG classifiers for transfer learning [7]. In this study, we propose a combination of transfer learning and an exponentially dilated causal convolutional network for automated comprehensive interpretation of the ECG.…”

Transfer learning for ECG classification

“…Here, we deal with ECG frames much longer than a heartbeat, which we use to classify heart arrhythmia. More recently, Strodthoff et al 28 used transfer learning on public ECG data sets to classify heart arrhythmia. Similar to their work, we finetune the pretrained networks to classify heart arrhythmia, however, we use a much larger upstream data set for pretraining and investigate several pretraining tasks.…”

On the Application of Convolutional Neural Networks for 12-lead ECG Multi-label Classification using Datasets from Multiple Centers

“…On top of the convolutional module a concat-pooling layer was used (concatenation of the output obtained with global average and max poolings) as was done in Strodthoff et al [10]. In addition, all the re-implemented CNN architectures shared the same dense module, implemented as a first fully-connected layer with 128 units followed by batch normalization, ReLU non-linearity and dropout ( = 0.5), and a second fully-connected layer with 27 units (output layer) activated with sigmoid functions.…”

“…This provided an end-to-end framework where the most relevant features are automatically learned directly from raw/lightly pre-processed data without separately perform feature extraction and classification. When applied to ECG signals [8][9][10], this enabled statements that resulted highly difficult to make even for cardiologists [11]. Among DNNs, recently Strodthoff et al [10] reported outstanding results using deep convolutional neural networks (CNNs), such as InceptionTime [12], ResNet [13] and XResNets [14], on a large public benchmark dataset.…”