Beat by Beat: Classifying Cardiac Arrhythmias with Recurrent Neural Networks

Introduction

Cardiac arrhythmias are a heterogenous group of conditions that is characterised by heart rhythms that do not follow a normal sinus pattern. One of the most common arrhythmias is atrial fibrillation (AF) with an age-dependant population prevalence of $2.3$ - $3.4\%$ . Due to the increased mortality associated with arrhythmias, receiving a timely diagnosis is of paramount importance for patients . To diagnose cardiac arrhythmias, medical professionals typically consider a patient’s electrocardiogram (ECG) as one of the primary factors . In the past, clinicians recorded these ECGs mainly using multi-lead clinical monitors or Holter devices. However, the recent advent of mobile cardiac event recorders has given patients the ability to remotely record short ECGs using devices with a single lead.

We propose a machine-learning approach based on recurrent neural networks (RNNs) to differentiate between various types of heart rhythms in this more challenging setting with just a single lead and short ECG record lengths. To ease learning of dependencies over the temporal dimension, we introduce a novel task formulation that harnesses the natural beat-wise segmentation of ECG signals. In addition to utilising several heartbeat features that have been shown to be highly discriminative in previous works, we also use stacked denoising autoencoders (SDAE) to capture differences in morphological structure. Furthermore, we extend our RNNs with a soft attention mechanism that enables us to reason about which ECG segments the RNNs prioritise for their decision making.

Methodology

Our cardiac rhythm classification pipeline consists of multiple stages (figure 1). The core idea of our setup is to extract a diverse set of features from the sequence of heartbeats in an ECG record to be used as input features to an ensemble of RNNs. We blend the individual models’ predictions into a per-class classification score using a multilayer perceptron (MLP) with a softmax output layer. The following paragraphs explain the stages shown in figure 1 in more detail.

ECG Dataset. We use the dataset of the PhysioNet Computing in Cardiology (CinC) 2017 challenge which contains 12,186 unique single-lead ECG records of varying length. Experts annotated each of these ECGs as being either a normal sinus rhythm, AF, an other arrhythmia or too noisy to classify. The challenge organisers keep 3,658 ($30\%$) of these ECG records private as a test set. Additionally, we hold out a non-stratified random subset of $20\%$ of the public dataset as a validation set. For some RNN configurations, we further augment the training data with labelled samples extracted from other PhysioNet databases in order to even out misbalanced class sizes in the training set. As an additional measure against the imbalanced class distribution of the dataset, we weight each training sample’s contribution to the loss function to be inversely proportional to its class’ prevalence in the overall dataset.

Normalisation. Prior to segmentation, we normalise the ECG recording to have a mean value of zero and a standard deviation of one. We do not apply any additional filters as all ECGs were bandpass-filtered by the recording device.

Segmentation. Following normalisation, we segment the ECG into a sequence of heartbeats. We decide to reformulate the given task of classifying arrhythmias as a sequence classification task over heartbeats rather than over raw ECG readings. The motivation behind the reformulation is that it significantly reduces the number of time steps through which the error signal of our RNNs has to propagate. On the training set, the reformulation reduces the mean number of time steps per ECG from $9000$ to just $33$. To perform the segmentation, we use a customised QRS detector based on Pan-Tompkin’s that identifies R-peaks in the ECG recording. We extend their algorithm by adapting the threshold with a moving average of the ECG signal to be more resilient against the commonly encountered short bursts of noise. For the purpose of this work, we define heartbeats using a symmetric fixed size window with a total length of $0.66$ seconds around R-peaks. We pass the extracted heartbeat sequence in its original order to the feature extraction stage.

Feature Extraction. We extract a diverse set of features from each heartbeat in an ECG recording. Specifically, we extract the time since the last heartbeat ($\delta$RR), the relative wavelet energy (RWE) over five frequency bands, the total wavelet energy (TWE) over those frequency bands, the R amplitude, the Q amplitude, QRS duration and wavelet entropy (WE). Previous works demonstrated the efficacy of all of these features in discriminating cardiac arrhythmias from normal heart rhythms . In addition to the aforementioned features, we also train two SDAEs on the heartbeats in an unsupervised manner with the goal of learning more nuanced differences in morphology of individual heartbeats. We train one SDAE on the extracted heartbeats of the training set and the other on their wavelet coefficients. We then use the encoding side of the SDAEs to extract low-dimensional embeddings of each heartbeat and each heartbeat’s wavelet coefficients to be used as additional input features. Finally, we concatenate all extracted features into a single feature vector per heartbeat and pass them to the level 1 models in original heartbeat sequence order.

Level 1 Models. We build an ensemble of level 1 models to classify the sequence of per-beat feature vectors. To increase the diversity within our ensemble, we train RNNs in various binary classification settings and with different hyperparameters. We use RNNs with $1$ - $5$ recurrent layers that consist of either Gated Recurrent Units (GRU) or Bidirectional Long Short-Term Memory (BLSTM) units , followed by an optional attention layer, $1$ - $2$ forward layers and a softmax output layer. Additionally, we infer a nonparametric Hidden Semi-Markov Model (HSMM) with $64$ initial states for each class in an unsupervised setting. In total, our ensemble of level 1 models consists of 15 RNNs and 4 HSMMs. We concatenate the ECG’s normalised log-likelihoods under the per-class HSMMs and the RNNs’ softmax outputs into a single prediction vector. We pass the prediction vector of the level 1 models to the level 2 blender model.

Level 2 Blender. We use blending to combine the predictions of our level 1 models and a set of ECG-wide features into a final per-class classification score. The additional features are the RWE and WE over the whole ECG and the absolute average deviation (AAD) of the WE and $\delta$RR of all beats. We employ a MLP with a softmax output layer as our level 2 blender model. In order to avoid overfitting to the training set, we train the MLP on the validation set.

Hyperparameter Selection. To select the hyperparameters of our level 1 RNNs, we performed a grid search on the range of - $75\%$ for the dropout and recurrent dropout percentages, $60$ - $512$ for the number of units per hidden layer and $1$ - $5$ for the number of recurrent layers. We found that RNNs trained with $35\%$ dropout, $65\%$ recurrent dropout, $80$ units per hidden layer and $5$ recurrent layers (plus an additional attention layer) achieve consistently strong results across multiple binary classification settings. For our level 2 blender model, we utilise Bayesian optimisation to select the number of layers, number of hidden units per layer, dropout and number of training epochs. We perform a 5-fold cross validation on the validation set to select the blender model’s hyperparameters.

Attention mechanisms have been shown to allow for greater interpretability of neural networks in a variety of tasks in computer vision and natural language processing . In this work, we apply soft attention over the heartbeats contained in ECG signals in order to gain a deeper understanding of the decision-making process of our RNNs. Consider the case of an RNN that is processing a sequence of $T$ heartbeats. The topmost recurrent layer outputs a hidden state $h_{t}$ at every time step $t\in[1,T]$ of the sequence. We extend some of our RNNs with additive soft attention over the hidden states $h_{t}$ to obtain a context vector $c$ that attenuates the most informative hidden states $h_{t}$ of a heartbeat sequence. Based on the definition in , we use the following set of equations:

Where equation (1) is a single-layer MLP with a weight matrix $W_{beat}$ and bias $b_{beat}$ to obtain $u_{t}$ as a hidden representation of $h_{t}$ . In equation (2), we calculate the attention factors $a_{t}$ for each heartbeat by computing a softmax over the dot-product similarities of every heartbeat’s $u_{t}$ to the heartbeat context vector $u_{beat}$. $u_{beat}$ corresponds to a hidden representation of the most informative heartbeat . We jointly optimise $W_{beat}$, $b_{beat}$ and $u_{beat}$ with the other RNN parameters during training. In figure 2, we showcase two examples of how qualitative analysis of the attention factors $a_{t}$ of equation (2) provides a deeper understanding of our RNNs’ decision making.

Related Work

Our work builds on a long history of research in detecting cardiac arrhythmias from ECG records by making use of features that have been shown to be highly discriminative in distinguishing certain arrhythmias from normal heart rhythms . Recently, Rajpurkar et al. proposed a 34-layer convolutional neural network (CNN) to reach cardiologist-level performance in classifying a large set of arrhythmias from mobile cardiac event recorder data . In contrast, we achieve state-of-the-art performance with significantly fewer trainable parameters by harnessing the natural heartbeat segmentation of ECGs and discriminative features from previous works. Additionally, we pay consideration to the fact that interpretability remains a challenge in applying machine learning to the medical domain by extending our models with an attention mechanism that enables medical professionals to reason about which heartbeats contributed most to the decision-making process of our RNNs.

Results and Conclusion

We present a machine-learning approach to distinguishing between multiple types of heart rhythms. Our approach utilises an ensemble of RNNs to jointly identify temporal and morphological patterns in segmented ECG recordings of any length. In detail, our approach reaches an average F1 score of 0.79 on the private test set of the PhysioNet CinC Challenge 2017 ($n=3,658$) with class-wise F1 scores of $0.90$, $0.79$ and $0.68$ for normal rhythms, AF and other arrhythmias, respectively. On top of its state-of-the-art performance, our approach maintains a high degree of interpretability through the use of a soft attention mechanism over heartbeats. In the spirit of open research, we make an implementation of our cardiac rhythm classification system available through the PhysioNet 2017 Open Source Challenge.

Future Work. Based on our discussions with a cardiologist, we hypothesise that the accuracy of our models could be further improved by incorporating contextual information, such as demographic information, data from other clinical assessments and behavioral aspects.

Acknowledgements

This work was partially funded by the Swiss National Science Foundation (SNSF) project No. 167302 within the National Research Program (NRP) $75$ “Big Data” and SNSF project No. 150640. We thank Prof. Dr. med. Firat Duru for providing valuable insights into the decision-making process of cardiologists.

References