Classiﬁcation of Cardiac Arrhythmias from Single Lead ECG

with a Convolutional Recurrent Neural Network

ome Van Zaen, Olivier Ch

etelat, Mathieu Lemay,

Enric M. Calvo and Ricard Delgado-Gonzalo

Swiss Center for Electronics and Microtechnology (CSEM), Rue Jaquet-Droz 1, Neuch

atel, Switzerland

Keywords:

ECG, Cardiac Arrhythmias, Neural Networks, Deep Learning, Wearable Sensors.

Abstract:

While most heart arrhythmias are not immediately harmful, they can lead to severe complications. In par-

ticular, atrial ﬁbrillation, the most common arrhythmia, is characterized by fast and irregular heart beats and

increases the risk of suffering a stroke. To detect such abnormal heart conditions, we propose a system com-

posed of two main parts: a smart vest with two cooperative sensors to collect ECG data and a neural network

architecture to classify heart rhythms. The smart vest uses two dry bi-electrodes to record a single lead ECG

signal. The biopotential signal is then streamed via a gateway to the cloud where a neural network detects and

classiﬁes the heart arrhythmias. We selected an architecture that combines convolutional and recurrent layers.

The convolutional layers extract relevant features from sliding windows of ECG and the recurrent layer aggre-

gates them for a ﬁnal softmax layer that performs the classiﬁcation. Our neural network achieves an accuracy

of 87.50% on the dataset of the challenge of Computing in Cardiology 2017.

1 INTRODUCTION

Heart arrhythmias are caused by irregular electrical

conduction in cardiac tissue. Atrial ﬁbrillation, which

affects 1–2% of the population (Camm et al., 2010),

is the most common one. Furthermore, its prevalence

increases with age, from <0.5% at 40–50 years to 5–

15% at 80 years. While not directly life-threatening,

it can lead to serious complications (January et al.,

2014). In particular, atrial ﬁbrillation is associated

with a 3–5 fold increased risk of stroke and a 2-fold

increased risk of mortality (Kannel et al., 1998). It

is also associated with a 3-fold risk of heart failure

(Wang et al., 2003). Typical symptoms include heart

palpitations, shortness of breath, and fainting. How-

ever, about one third of the cases are asymptomatic

which prevents early diagnosis. This, in turn, pre-

cludes early therapies which might protect the patient

from the consequences of atrial ﬁbrillation but also

from its progression. Indeed, atrial ﬁbrillation causes

electrical and structural remodeling of the atria which

facilitates its further development (Frick et al., 2001;

Nattel et al., 2008).

The gold standard for diagnosing atrial ﬁbrillation

and other heart arrhythmias is the 12-lead ECG. A

trained electrophysiologist can select the most appro-

priate treatment after reviewing ECG signals and the

patient history. This is, however, a time-consuming

task, especially for long recordings such as the ones

collected with Holter monitors. To alleviate this

task, several approaches have been proposed to de-

tect arrhythmias from ECG signals (Owis et al., 2002;

De Chazal et al., 2004). Even without perfect de-

tection accuracy, these approaches are useful as they

facilitate reviewing ECG by selecting relevant signal

excerpts.

Recently, neural networks have shown impressive

performance in various classiﬁcation and regression

tasks. Image processing was the ﬁrst ﬁeld where

deep networks surpassed existing approaches by a

large margin (Krizhevsky et al., 2012). Since then,

they have been extensively applied to ﬁelds previ-

ously dominated by signal processing. In particular,

several architectures have been proposed to detect and

classify heart arrhythmias from ECG signals.

In the context of the challenge of Computing in

Cardiology 2017 (Clifford et al., 2017), a few neu-

ral network architectures were proposed to classify

single lead ECG signals into one of the following

classes: normal sinus rhythm, atrial ﬁbrillation, other

rhythm, and noise. One of these architectures uses

logarithmic spectrograms computed over sliding win-

dows of ECG as input to two-dimensional convolu-

tional layers (Zihlmann et al., 2017). Aggregation

Van Zaen, J., Chételat, O., Lemay, M., Calvo, E. and Delgado-Gonzalo, R.

Classiﬁcation of Cardiac Arrhythmias from Single Lead ECG with a Convolutional Recurrent Neural Network.

DOI: 10.5220/0007347900330041

In Proceedings of the 12th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2019), pages 33-41

ISBN: 978-989-758-353-7

of successive windows is done with either temporal

averaging or a recurrent layer. However, the convo-

lutional and recurrent layers were trained separately

due to convergence issues. A similar approach used

a 16-layer convolutional neural network with skip

connections to classify arrhythmias from ECG sig-

nals (Xiong et al., 2017). Each layer is composed of

batch normalization, ReLU activation, dropout, one-

dimensional convolution, and averaging pooling.

Recently, a convolutional neural network was

shown to reach cardiologist-level arrhythmia detec-

tion (Rajpurkar et al., 2017). This 34-layer network

takes advantage of a very large dataset of 64,121

ECG signals, recorded from 29,163 patients, to rec-

ognize 12 different heart arrhythmias including atrial

ﬁbrillation, atrial ﬂutter, and ventricular tachycar-

dia. Another approach applied convolutional neural

networks for detecting atrial ﬁbrillation from time-

frequency representations of ECG signals (Xia et al.,

2018). Two methods for computing these represen-

tations were compared: the short-time Fourier trans-

form or the stationary wavelet transform. In this case,

the neural network using coefﬁcients from the second

transform yielded better classiﬁcation accuracy.

Neural networks have thus shown promising re-

sults for the detection of abnormal cardiac rhythms.

Furthermore, as mentioned previously, it is of the ut-

most importance to detect arrhythmias as early as pos-

sible to improve treatment outcome. To tackle this is-

sue, we developed a system composed of two main el-

ements: a smart vest to record ECG signals and an al-

gorithm to detect and classify arrhythmias. The smart

vest includes two cooperative sensors to record a sin-

gle lead ECG signal and stream by Bluetooth the col-

lected data to a gateway. This gateway then forwards

the ECG signal to the cloud where it is processed by a

neural network in order to detect abnormal rhythms.

This article is structured as follows. First, the

dataset of ECG signals, our neural network architec-

ture, and the monitoring system are described in Sec-

tion 2. Then, the results are presented in Section 3

and discussed in Section 4. Finally, this article ends

with a short conclusion in Section 5.

2 MATERIALS AND METHODS

2.1 Dataset

We used the dataset from the challenge of Computing

in Cardiology 2017 to train a neural network to clas-

sify cardiac arrhythmias. This dataset includes 8528

single lead ECG signals recorded with an AliveCor

device. The signals are sampled at 300 Hz and have

Table 1: Number of signals and mean duration for each

class.

Class Count Proportion Mean duration [s]

Normal rhythm 5076 59.52% 32.11

Atrial ﬁbrillation 758 8.89% 32.34

Other rhythm 2415 28.32% 34.30

Noise 279 3.27% 24.38

Total 8528 100% 32.50

durations ranging from 9 to 60 seconds. Each signal

was acquired when the subject held each one of the

two electrodes in each hand resulting in a lead I (left

arm – right arm) ECG. As the device has no speciﬁc

orientation, many signals are inverted (right arm – left

arm).

All ECG signals are labeled with one of the fol-

lowing four classes: normal sinus rhythm, atrial ﬁb-

rillation, other rhythm, and noise. The proportion of

each class in the dataset varies from 3.27% for noise

to 59.52% for normal rhythm. The full breakdown of

all classes is reported in Table 1.

The entries of the challenge of Computing in Car-

diology 2017 were ranked according to the following

score evaluated on a private test set:

CinC

+ F

(1)

where F

, F

, and F

denote the F

scores for nor-

mal rhythm, atrial ﬁbrillation, and other rhythm. The

four winners (Teijeiro et al., 2017; Datta et al., 2017;

Zabihi et al., 2017; Hong et al., 2017) reached a score

of 0.83.

Several aspects of this dataset are challenging for

arrhythmia classiﬁcation. First, many signals are in-

verted as mentioned previously. Second, the classes

are not balanced. There are very few signals labeled

atrial ﬁbrillation or noise compared to the ones la-

beled normal rhythm. Furthermore, the durations of

the recordings are also different. They vary from 9

to 60 seconds. Figure 1 illustrates this issue. Most

ECG signals last 30 seconds but a signiﬁcant portion

has shorter or longer durations. In addition, labeling

is relatively coarse as a single label is associated with

each ECG signal. In some cases, several labels could

be used for the same signal. Finally, the ECG quality

of a non-negligible part of the records is rather poor.

Four examples of signals are shown in Figure 2. The

ﬁrst two signals are labeled normal rhythm and atrial

ﬁbrillation and have good overall quality. The third

example is a normal rhythm record with acceptable

quality except for a short segment of noise. The last

ECG signal is labeled as atrial ﬁbrillation but has very

BIOSIGNALS 2019 - 12th International Conference on Bio-inspired Systems and Signal Processing

Figure 1: Record durations sorted in increasing order.

Figure 2: Examples of ECG records: (A) normal rhythm

from record A00001, (B) atrial ﬁbrillation from record

A00004, (C) normal rhythm from record A00002, (D) atrial

ﬁbrillation from record A00015.

poor quality due to large shifts of the baseline. It also

illustrates that all signals do not share the same dura-

tion.

The test set used during the challenge of Com-

puting in Cardiology was not yet publicly released.

Therefore, we split the dataset (which was originally

intended only for training) into a training set of 7000

records and a test set with the remaining 1528 records

while approximately preserving class repartition.

We applied the following pre-processing steps to

the dataset. First, we ﬁltered the ECG signals with a

digital Butterworth band-pass ﬁlter between 0.5 and

40 Hz. The ﬁlter is applied in both forward and back-

ward directions to avoid distortions. The analog ﬁl-

ter used by the recording device has similar cutoff

frequencies but still leaked some components outside

the pass-band. Then, we downsampled the signals to

200 Hz to reduce the number of samples. Finally, the

signals are scaled by the mean of the standard devi-

ations of all signals from the training set. Scaling is

helpful to accelerate training (LeCun et al., 2012).

2.2 Network Architecture

An approach to handle signals with different dura-

tions is to extract windows with the same length. The

label of a signal is then used for all included windows.

Fixed size inputs would make possible to apply a con-

volutional neural network to learn features useful to

classify arrhythmias. However, this approach is sub-

optimal as illustrated in Figure 2 where a signal la-

beled as normal rhythm includes a segment of noise.

Furthermore, all signals would need to be truncated to

the same length. This would lead to a large data loss

due to the considerable variations in signal durations.

A more appropriate approach is to use a recur-

rent neural network which is well-suited to process

sequences with varying lengths. While such a neural

network can, by design, remember past values over

long time intervals, they are not as efﬁcient as con-

volutional neural networks for learning complex fea-

tures.

After reviewing the advantages and drawbacks of

these two approaches, we chose to combine them and

build a neural network that includes convolutional

and recurrent layers. Speciﬁcally, each ECG signal

was divided in sliding windows of 512 samples with

50% overlap. This corresponds to a window dura-

tion slightly above 2.5 seconds. The number of win-

dows extracted from each signal depends on its du-

ration. Seven convolutional layers were applied to

all windows of a signal. Each convolutional layer

is composed of a 1D convolution and a max pool-

ing operation. The convolution uses a kernel of size

5, zero padding, and a ReLU activation (Hahnloser

et al., 2000). The pool size for max pooling was set

to 2. The ﬁrst convolutional layer has 8 output chan-

nels (from the single channel windows). Then, each

following layers double the number of channels while

max pooling halves the number of samples. After the

convolutional layers, a global averaging pooling layer

was applied. This results in 512 features for each in-

put window. The features were then processed with

a long short-term memory (LSTM) layer (Hochreiter

and Schmidhuber, 1997) with 128 units. Finally, a

softmax layer outputs the probability of each class for

the input ECG windows. This results in a neural net-

work with 1,203,364 trainable parameters including

874,656 for the convolutional part, 328,192 for the re-

Classiﬁcation of Cardiac Arrhythmias from Single Lead ECG with a Convolutional Recurrent Neural Network

Table 2: Neural network architecture. The output size is

given as N

× N

where N

denotes the variable num-

ber of windows, N

the number of samples, and N

the num-

ber of channels.

Layer Output size

Input windows N

× 512 ×1

Convolutional layer 1 N

× 256 × 8

Convolutional layer 2 N

× 128 × 16

Convolutional layer 3 N

× 64 × 32

Convolutional layer 4 N

× 32 × 64

Convolutional layer 5 N

× 16 × 128

Convolutional layer 6 N

× 8 × 256

Convolutional layer 7 N

× 4 × 512

Global average pooling N

× 512

LSTM layer 128

Softmax layer 4

current part, and 516 for the ﬁnal softmax layer. The

complete network architecture is summarized in Ta-

ble 2.

2.3 Data Augmentation

As the dataset is relatively small for ﬁtting a neural

network, we applied different strategies to syntheti-

cally augment the number of ECG signals available

during training. The ﬁrst strategy is to simply ﬂip the

sign of each signal with probability 0.5. We found

it easier to let the neural network learn to take into

account inverted ECG signals instead of applying a

rectifying step during pre-processing.

Furthermore, when extracting sliding windows, it

is not possible to use all samples for the large ma-

jority of ECG signals. Indeed, the maximum number

of sliding windows N

in a signal with N samples is

given by



N − 512

256



+ 1

assuming N ≥ 512. In the previous expression, b·c

denotes the ﬂoor function. We took advantage of this

fact to place the ﬁrst window at a random offset from

the start of the signal. This random offset is drawn

uniformly from

{0, 1, . . . ,N − (N

− 1) · 256 − 512}

for each signal at each epoch. The main idea behind

this strategy is to prevent the neural network from

learning the exact positions of the QRS complexes in

the training set. However, we always used the max-

imum number of sliding windows possible for each

signal to avoid wasting ECG samples.

The third strategy we applied is to resample each

signal at each epoch with probability 0.8 in order to

simulate slightly slower or faster heart rate and thus

help the neural network to reach better generaliza-

tion performance. Naturally, the resampling opera-

tion should not change the heart rate too much. Oth-

erwise, there is a risk to confuse a cardiac rhythm

for another one. Therefore, if a signal needs resam-

pling, its length is changed by a proportion sampled

uniformly between −5% and +5%.

2.4 Training

We implemented our neural network and strategies for

data augmentation in Python with the Keras library

(Chollet et al., 2015). We trained the neural network

for 200 epochs by minimizing the cross-entropy with

the Adam algorithm (Kingma and Ba, 2014). We set

the initial learning rate to 0.001. The learning rate

was divided by two if the cross-entropy evaluated on

the test set did not decrease for 5 consecutive epochs

with a lower limit at 10

−5

We used a batch size of 50 signals. As a batch

must include the same number of sliding windows for

each signal, we applied zero-padding. Speciﬁcally,

too short signals were prepended with all-zero win-

dows. To limit zero-padding as much as possible we

sorted the signals by duration and grouped them in

batches of similar lengths. This resulted in batches

with varying numbers of windows.

The LSTM layer was regularized by applying

dropout with a rate of 0.5 for both the input and re-

current parts (Srivastava et al., 2014; Gal and Ghahra-

mani, 2016). We monitored the accuracy on the test

set and selected the weights at the best epoch as the

ﬁnal parameters of the neural network.

2.5 Monitoring System

The smart vest used to monitor ECG includes two

cooperative sensors illustrated in Figure 3. These

sensors use dry stainless steel bi-electrodes. This

technology, which was validated in a previous study,

yields high quality measurements of ECG and bio-

impedance signals, even in motion. No wetting of the

electrode is required. Moreover, the electrical connec-

tion linking both sensors does not have to be shielded,

nor insulated, which makes its integration in garment

easier and cheaper (a conductive fabric is sufﬁcient).

The length of the connection is not limited to very

short distances and is therefore placed in the back so

BIOSIGNALS 2019 - 12th International Conference on Bio-inspired Systems and Signal Processing

Figure 3: Cooperative sensors included in the smart vest for

monitoring ECG.

Figure 4: Smart vest during the validation protocol.

as to have space for a central zipper in the vest, which

makes donning and dofﬁng easy, even for the elderly.

The two watertight sensors are clipped in the vest and

can be removed for washing and recharging. Opera-

tion is simpliﬁed to its maximum: the sensors auto-

matically switch to record mode as soon as they are

applied on the skin and return to standby mode when

removed.

This smart vest which was originally developed

for athletes also monitors the following biomedi-

cal signals: heart rate, transthoracic impedance, res-

piration rate, skin temperature, activity class (rest-

ing, walking, running), and posture (lying, stand-

ing/sitting). A quality index is associated with each

signal so that the reliability of the measurements can

be easily assessed. However, these additional signals

and the quality indices were not used in the present

study. Figure 4 shows the smart vest during the vali-

dation protocol.

Another important feature of the smart vest con-

sists in its capability to store all the recorded signals

locally as well as synchronize them in the cloud. This

is achieved by means of an accompanying gateway

that acts as a relay for the streamed data. The gate-

way is robust to poor network connectivity and up-

loads the newly available data to the cloud as it is

being streamed while an active connection is avail-

able. In our case, the gateway is composed of a simple

Raspberry Pi that connects to the smart vest via Blue-

tooth. The biomedical data is then collected, stored

locally in a compact format and relayed via telemetry

messages simultaneously to all the clients connected

to this gateway as well as a private cloud. The mes-

saging protocol chosen for this application is MQTT

(Message Queuing Telemetry Transport). This pro-

tocol, which uses the publish/subscribe paradigm, is

one of the most widely employed telemetry protocols

in IoT and real-time streaming. Once the ECG signal

is in the cloud, we can leverage the powerful comput-

ing capabilities and detect heart arrhythmias with our

neural network.

The ECG signals recorded with the smart vest de-

vice were pre-processed similarly to the data from the

challenge of Computing in Cardiology 2017. First,

the same band-pass ﬁlter between 0.5 and 40 Hz was

used to remove the baseline and high-frequency noise.

Then, the signals were resampled at 200 Hz. Finally,

we scaled each signal with its standard deviation. In-

deed, we could not use the scaling factor computed

on the training set as the two types of ECG signals

did not have the same range of values.

After pre-processing the ECG signals, we ex-

tracted sliding windows of 512 samples with 50%

overlap. We used groups of 25 such windows as in-

put to the neural network. We selected this speciﬁc

number of windows as it corresponds to segments of

approximately 33 seconds which is close to the me-

dian length of the signals used for training and testing

the model. Applying the neural network resulted in a

rhythm prediction for each group of 25 windows.

3 RESULTS

We evaluated three conﬁgurations. The ﬁrst one used

the architecture described in Table 2 except that it

included only 6 convolutional layers. In addition,

all data augmentation strategies were applied during

training except resampling. The second conﬁguration

used the full architecture with 7 convolutional layers

but again without resampling. Finally, the last con-

ﬁguration was identical to the second one but with

resampling for further data augmentation. Figure 5

shows the evolution of the cross-entropy loss and the

accuracy during training for these three network con-

ﬁgurations. An additional convolutional layer helped

to increase the accuracy and reduce the loss. Resam-

pling the signals also slightly improved performance.

Despite our efforts, we could not completely elimi-

nate over-ﬁtting as shown by the performance gap be-

tween training and test sets. Indeed, additional regu-

larization only decreased the performance of the net-

work. However, it is worth mentioning that resam-

https://mqtt.org

Classiﬁcation of Cardiac Arrhythmias from Single Lead ECG with a Convolutional Recurrent Neural Network

Figure 5: Cross-entropy loss (top) and accuracy (bottom)

evaluated during training for a network with 6 convolutional

layers without resampling (blue), a network with 7 con-

volutional layers without resampling (orange), and a net-

work with 7 convolutional layers with resampling (green).

Dashed lines denote results obtained on the training set and

solid lines results obtained on the test set.

pling helped to reduce over-ﬁtting. The neural net-

work with the third conﬁguration reached an accuracy

of 87.37% on the test set after 47 epochs. At the same

epoch, the accuracy on the training set was 90.90%.

After selecting the best neural network, we evalu-

ated it without zero-padding by selecting a batch size

of 1. In this case, the accuracy was 92.64% on the

training set and 87.50% on the test set. Sensitivity,

speciﬁcity, and F

score values are reported in Ta-

ble 3 for all classes. Unsurprisingly, the best F

score

was obtained for the class with the most samples (nor-

mal rhythm) and the worst one for the class with the

least samples (noise). Interestingly, the speciﬁcity for

atrial ﬁbrillation was relatively high at 0.9784 while

the sensitivity was lower at 0.8382. Thus, the num-

ber of false positive is more larger than the number of

false negative which is an important property of the

model. Indeed, it missed only a few atrial ﬁbrillation

cases while false detections can always be disproved

with additional analyses such as a full 12-lead ECG.

Furthermore, when evaluated in terms of the score

used during the challenge of Computing in Cardiol-

ogy 2017 (1), the neural network yielded 0.9156 and

0.8495 on the training and test sets. This is similar

to the best score obtained by the winners of the chal-

lenge. However, we could only evaluate our neural

network on 1528 ECG signals from the original train-

ing set since the ofﬁcial test set was not made public

yet. Therefore, it is difﬁcult to compare the perfor-

Table 3: Performance metrics for the best neural network.

Class Metric Training set Test set

Normal rhythm

Sensitivity 0.9707 0.9263

Speciﬁcity 0.8955 0.8853

score 0.9509 0.9243

Atrial ﬁbrillation

Sensitivity 0.9084 0.8382

Speciﬁcity 0.9942 0.9784

score 0.9232 0.8143

Other rhythm

Sensitivity 0.8380 0.7921

Speciﬁcity 0.9675 0.9352

score 0.8728 0.8099

Noise

Sensitivity 0.9345 0.7600

Speciﬁcity 0.9972 0.9871

score 0.9264 0.7103

mance of our approach with the winning entries.

After training and evaluation on the dataset of

the challenge of Computing in Cardiology 2017, we

applied the neural network to a few ECG signals

recorded with the smart vest. We also extracted the

times between consecutive R-waves from the ECG

signals (a.k.a. RR intervals) to facilitate visualization

of the heart rate. An example is shown in Figure 6.

In this case, the subject had a normal sinus rhythm

with two successive premature ventricular contrac-

tions characterized a very short RR interval followed

by a long recovery RR interval. These premature con-

tractions are clearly identiﬁable especially compared

to the stable RR intervals of a normal rhythm. The

neural network classiﬁed the this ECG segment as

normal rhythm except for the groups of windows in-

cluding the premature contractions which are classi-

ﬁed as other rhythm.

Another example of classiﬁcation is shown in Fig-

ure 7. In this case, no ECG signal was collected dur-

ing a few minutes due to poor skin-electrode contact

caused by motion. The neural network identiﬁed this

segment without signal as noise. Furthermore, two

segments preceding signal loss were classiﬁed as nor-

mal rhythm and other rhythm. The RR intervals were

stable in the ﬁrst segment and there were a few prema-

ture ventricular contractions in the second one. After

the ECG signal was recovered, the RR intervals var-

ied widely and the neural network classiﬁed the data

as other rhythm.

BIOSIGNALS 2019 - 12th International Conference on Bio-inspired Systems and Signal Processing

Figure 6: Cardiac rhythm classiﬁcation from an ECG signal recorded with the smart vest. The ECG signal (top) includes two

premature ventricular contractions which are visible in the RR intervals (middle). The neural network correctly identiﬁes this

segment of the signal as reﬂected by the class probabilities (bottom).

Figure 7: Cardiac rhythm classiﬁcation from an ECG signal recorded with the smart vest. The ECG signal (top) includes a

segment with no R waves due to poor skin-electrode contact. Consequently, the RR intervals (middle) could not be extracted.

The class probabilities (bottom) computed by the neural network are valid as the segment is labeled as noise.

Classiﬁcation of Cardiac Arrhythmias from Single Lead ECG with a Convolutional Recurrent Neural Network

4 DISCUSSION

The neural network architecture we developed

achieved a classiﬁcation performance similar to the

winners of the challenge of Computing in Cardiology

2017. Although we could not evaluate our approach

on the same data since the true test set was not re-

leased yet, the results are promising. In particular,

the speciﬁcity for atrial ﬁbrillation shows that only a

limited number cases are not detected. We could take

into account signals of different lengths by combin-

ing convolutional and recurrent layers. Indeed, the

convolutional layers extract features relevant for ar-

rhythmia classiﬁcation from sliding windows of raw

ECG signal. With this approach, there is no need to

pre-process the data to extract spectrogram or wavelet

transform for instance. Furthermore, no feature engi-

neering is required as the neural network learns dur-

ing the training phase to extract high-level features

useful for classiﬁcation. The only pre-processing we

used is band-pass ﬁltering as well as scaling to accel-

erate training. We also used strategies for data aug-

mentation to reduce over-ﬁtting and improve the gen-

eralization performance of the network. In particu-

lar, randomly ﬂipping the sign of ECG signals forced

the neural network to learn to take into account both

regular and inverted waveforms. Without this simple

strategy, we would have needed to develop a robust

method to detect signals that had to be rectiﬁed.

In addition, we have shown that it is possible to

apply a neural network trained on a generic dataset

to ECG signals recorded by the smart vest with min-

imal changes. Indeed, there was no need to adapt

the network architecture. The method to standardize

the ECG signals was the only element that required a

modiﬁcation. Of course, if the two measurement sys-

tems, the AliveCor device and the smart vest, did not

record the same ECG lead, additional changes would

be needed. However, it is difﬁcult to determine the

extend of these changes without evaluating the neu-

ral network on a dataset of ECG signals from another

lead.

Taken together, these results demonstrate that our

system could be used to monitor the cardiac activity

of a subject for detecting abnormal rhythms over long

periods of time. Indeed, the smart vest is more com-

fortable to wear than a traditional Holter and it can

still collect high quality ECG signals since it was ini-

tially developed for athletes. After transmitting the

data to the cloud, our neural network can quickly

classify abnormal rhythms. While the accuracy of

our system is not perfect, it can still help to reduce

the time spent reviewing ECG by selecting segments

with potential abnormalities that require additional at-

tention. These segments can then be analyzed by a

trained specialist. If needed, a full 12-lead ECG can

be performed to reﬁne or conﬁrm the diagnosis.

5 CONCLUSION

We presented a system composed of a smart vest to

record a single lead ECG signal and a neural network

for detection and classiﬁcation of heart arrhythmias.

We plan to aggregate several databases of ECG sig-

nals with rhythm annotations to extend the types of ar-

rhythmias that our algorithm can detect and improve

its accuracy. We will also investigate whether adding

skip connections in our network architecture further

improves classiﬁcation performance.

ACKNOWLEDGEMENTS

We would like to thank the anonymous reviewers for

their valuable suggestions and comments.

REFERENCES

Camm, A. J. et al. (2010). Guidelines for the manage-

ment of atrial ﬁbrillation. European Heart Journal,

31(19):2369–2429.

Chollet, F. et al. (2015). Keras. https://keras.io.

Clifford, G. D., Liu, C., Moody, B., Lehman, L.-w. H.,

Silva, I., Li, Q., Johnson, A., and Mark, R. G. (2017).

AF classiﬁcation from a short single lead ECG record-

ing: The PhysioNet/Computing in Cardiology Chal-

lenge 2017. Proceedings of Computing in Cardiology,

44:1.

Datta, S., Puri, C., Mukherjee, A., Banerjee, R., Choudhury,

A. D., Singh, R., Ukil, A., Bandyopadhyay, S., Pal, A.,

and Khandelwal, S. (2017). Identifying normal, AF

and other abnormal ECG rhythms using a cascaded

binary classiﬁer. In 2017 Computing in Cardiology

(CinC), pages 1–4.

De Chazal, P., O’Dwyer, M., and Reilly, R. B. (2004). Auto-

matic classiﬁcation of heartbeats using ECG morphol-

ogy and heartbeat interval features. IEEE Transac-

tions on Biomedical Engineering, 51(7):1196–1206.

Frick, M., Frykman, V., Jensen-Urstad, M., and

Ostergren,

J. (2001). Factors predicting success rate and re-

currence of atrial ﬁbrillation after ﬁrst electrical car-

dioversion in patients with persistent atrial ﬁbrillation.

Clinical Cardiology, 24(3):238–244.

Gal, Y. and Ghahramani, Z. (2016). A theoretically

grounded application of dropout in recurrent neural

networks. In Advances in neural information process-

ing systems, pages 1019–1027.

BIOSIGNALS 2019 - 12th International Conference on Bio-inspired Systems and Signal Processing

Hahnloser, R. H., Sarpeshkar, R., Mahowald, M. A., Dou-

glas, R. J., and Seung, H. S. (2000). Digital selec-

tion and analogue ampliﬁcation coexist in a cortex-

inspired silicon circuit. Nature, 405(6789):947.

Hochreiter, S. and Schmidhuber, J. (1997). Long short-term

memory. Neural Computation, 9(8):1735–1780.

Hong, S., Wu, M., Zhou, Y., Wang, Q., Shang, J., Li, H., and

Xie, J. (2017). ENCASE: An ENsemble ClASsiﬁEr

for ECG classiﬁcation using expert features and deep

neural networks. In 2017 Computing in Cardiology

(CinC), pages 1–4.

January, C. T., Wann, L. S., Alpert, J. S., Calkins, H., Cigar-

roa, J. E., Cleveland, J. C., Conti, J. B., Ellinor, P. T.,

Ezekowitz, M. D., Field, M. E., Murray, K. T., Sacco,

R. L., Stevenson, W. G., Tchou, P. J., Tracy, C. M., and

Yancy, C. W. (2014). 2014 AHA/ACC/HRS guide-

line for the management of patients with atrial ﬁbril-

lation: a report of the American College of Cardiol-

ogy/American Heart Association task force on prac-

tice guidelines and the Heart Rhythm Society. Journal

of the American College of Cardiology, 64(21):e1–

e76.

Kannel, W. B., Wolf, P. A., Benjamin, E. J., and Levy, D.

(1998). Prevalence, incidence, prognosis, and predis-

posing conditions for atrial ﬁbrillation: population-

based estimates. The American journal of cardiology,

82(7):2N–9N.

Kingma, D. P. and Ba, J. (2014). Adam: A

method for stochastic optimization. arXiv preprint

arXiv:1412.6980.

Krizhevsky, A., Sutskever, I., and Hinton, G. E. (2012). Im-

ageNet classiﬁcation with deep convolutional neural

networks. In Advances in Neural Information Pro-

cessing Systems 25, pages 1097–1105.

LeCun, Y. A., Bottou, L., Orr, G. B., and M

uller, K.-

R. (2012). Efﬁcient backprop. In Neural networks:

Tricks of the trade, pages 9–48. Springer.

Nattel, S., Burstein, B., and Dobrev, D. (2008). Atrial re-

modeling and atrial ﬁbrillation: mechanisms and im-

plications. Circulation: Arrhythmia and Electrophys-

iology, 1(1):62–73.

Owis, M. I., Abou-Zied, A. H., Youssef, A.-B., and Kadah,

Y. M. (2002). Study of features based on nonlinear

dynamical modeling in ECG arrhythmia detection and

classiﬁcation. IEEE Transactions on Biomedical En-

gineering, 49(7):733–736.

Rajpurkar, P., Hannun, A. Y., Haghpanahi, M., Bourn, C.,

and Ng, A. Y. (2017). Cardiologist-level arrhythmia

detection with convolutional neural networks. arXiv

preprint arXiv:1707.01836.

Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I.,

and Salakhutdinov, R. (2014). Dropout: a simple way

to prevent neural networks from overﬁtting. The Jour-

nal of Machine Learning Research, 15(1):1929–1958.

Teijeiro, T., Garc

ıa, C. A., Castro, D., and F

elix, P. (2017).

Arrhythmia classiﬁcation from the abductive interpre-

tation of short single-lead ECG records. In 2017 Com-

puting in Cardiology (CinC), pages 1–4.

Wang, T. J., Larson, M. G., Levy, D., Vasan, R. S., Leip,

E. P., Wolf, P. A., D’Agostino, R. B., Murabito, J. M.,

Kannel, W. B., and Benjamin, E. J. (2003). Temporal

relations of atrial ﬁbrillation and congestive heart fail-

ure and their joint inﬂuence on mortality: the Fram-

ingham heart study. Circulation, 107(23):2920–2925.

Xia, Y., Wulan, N., Wang, K., and Zhang, H. (2018). De-

tecting atrial ﬁbrillation by deep convolutional neu-

ral networks. Computers in Biology and Medicine,

93:84–92.

Xiong, Z., Stiles, M. K., and Zhao, J. (2017). Robust ECG

signal classiﬁcation for detection of atrial ﬁbrillation

using a novel neural network. In 2017 Computing in

Cardiology (CinC), pages 1–4.

Zabihi, M., Rad, A. B., Katsaggelos, A. K., Kiranyaz, S.,

Narkilahti, S., and Gabbouj, M. (2017). Detection

of atrial ﬁbrillation in ECG hand-held devices using

a random forest classiﬁer. In 2017 Computing in Car-

diology (CinC), pages 1–4.

Zihlmann, M., Perekrestenko, D., and Tschannen, M.

(2017). Convolutional recurrent neural networks for

electrocardiogram classiﬁcation. In 2017 Computing

in Cardiology (CinC), pages 1–4.

Classiﬁcation of Cardiac Arrhythmias from Single Lead ECG with a Convolutional Recurrent Neural Network