Beat by Beat: Classifying Cardiac Arrhythmias with Recurrent Neural Networks

Beat by Beat: Classifying Cardiac Arrhythmias with Recurrent Neural Networks. Schwab, P., Scebba, G., C., Zhang, J., Delai, M., & Karlen, W. Computing in Cardiology (CinC), 44:1-4, 9, 2017.

Website doi abstract bibtex

INTRODUCTION: Previous work on detecting arrhythmias in electrocardiogram (ECG) records has predominantly focused on identifying atrial fibrillation (AF) in data obtained from clinical settings or Holter devices, where long-term recordings with multiple leads are the norm. However, the advent of mobile cardiac event recorders increased the importance of being able to differentiate between multiple types of rhythms in noisy short-term recordings with just a single lead. We propose a machine-learning architecture to learn the temporal and morphological patterns of various types of rhythms in order to perform multiclass classification under these more challenging conditions. METHODS: We segment the input ECG signal with a QRS detector into individual heartbeats. From each heartbeat, we extract - among others - morphological features with the encoding side of a stacked denoising autoencoder that was trained in an unsupervised manner. The extracted features are passed in original heartbeat order as input sequences to an ensemble of recurrent neural networks (RNNs). The RNNs were trained on different features, random overlapping subsets of the training data and in various one-versus-all setups in order to increase the model diversity within the ensemble. We blend the individual RNNs’ predictions into a final classification solution using a multilayer perceptron (MLP) that was trained on held-out data. RESULTS: Our best ensemble at time of writing achieves an average F1-score over all classes of 0.78 (F1,normal=0.88, F1,af=0.75, F1,other=0.72, F1,noisy=0.78) on an out-of-sample test set (342 samples) and an average F1-score over all classes of 0.65 (F1,normal=0.82, F1,af=0.77, F1,other=0.64, F1,noisy=0.36) on the private test set for phase 1 of the PhysioNet 2017 challenge. CONCLUSION: Deep recurrent models enable our ensemble to differentiate between multiple types of heart rhythms by identifying temporal and morphological patterns in segmented ECG recordings of any length.

@article{
 title = {Beat by Beat: Classifying Cardiac Arrhythmias with Recurrent Neural Networks},
 type = {article},
 year = {2017},
 pages = {1-4},
 volume = {44},
 websites = {http://www.cinc.org/archives/2017/pdf/363-223.pdf},
 month = {9},
 day = {14},
 city = {Rennes, F},
 id = {39dd8240-969e-3db5-b8b4-afc49667129d},
 created = {2020-12-13T11:41:09.000Z},
 file_attached = {true},
 profile_id = {6d353feb-efe4-367e-84a2-0815eb9ca878},
 last_modified = {2022-09-04T18:12:03.468Z},
 read = {true},
 starred = {false},
 authored = {true},
 confirmed = {true},
 hidden = {false},
 citation_key = {Schwab2017},
 folder_uuids = {f1f67efc-95a7-4f1a-b181-c3670c667a34,4afa922c-d8d6-102e-ac9a-0024e85ead87,0801d9e0-d1ec-46e2-803d-c74946b43a02,d9198259-8733-497d-ab87-d2a9518e0d30},
 private_publication = {false},
 abstract = {INTRODUCTION: Previous work on detecting arrhythmias in electrocardiogram (ECG) records has predominantly focused on identifying atrial fibrillation (AF) in data obtained from clinical settings or Holter devices, where long-term recordings with multiple leads are the norm. However, the advent of mobile cardiac event recorders increased the importance of being able to differentiate between multiple types of rhythms in noisy short-term recordings with just a single lead. We propose a machine-learning architecture to learn the temporal and morphological patterns of various types of rhythms in order to perform multiclass classification under these more challenging conditions. METHODS: We segment the input ECG signal with a QRS detector into individual heartbeats. From each heartbeat, we extract - among others - morphological features with the encoding side of a stacked denoising autoencoder that was trained in an unsupervised manner. The extracted features are passed in original heartbeat order as input sequences to an ensemble of recurrent neural networks (RNNs). The RNNs were trained on different features, random overlapping subsets of the training data and in various one-versus-all setups in order to increase the model diversity within the ensemble. We blend the individual RNNs’ predictions into a final classification solution using a multilayer perceptron (MLP) that was trained on held-out data. RESULTS: Our best ensemble at time of writing achieves an average F1-score over all classes of 0.78 (F1,normal=0.88, F1,af=0.75, F1,other=0.72, F1,noisy=0.78) on an out-of-sample test set (342 samples) and an average F1-score over all classes of 0.65 (F1,normal=0.82, F1,af=0.77, F1,other=0.64, F1,noisy=0.36) on the private test set for phase 1 of the PhysioNet 2017 challenge. CONCLUSION: Deep recurrent models enable our ensemble to differentiate between multiple types of heart rhythms by identifying temporal and morphological patterns in segmented ECG recordings of any length.},
 bibtype = {article},
 author = {Schwab, Patrick and Scebba, Gaetano Claudio and Zhang, Jia and Delai, Marco and Karlen, Walter},
 doi = {10.22489/CinC.2017.363-223},
 journal = {Computing in Cardiology (CinC)}
}

Downloads: 0