84 Design Studies and Intelligence Engineering

L.C. Jain et al. (Eds.)

© 2023 The authors and IOS Press.
This article is published online with Open Access by IOS Press and distributed under the terms
of the Creative Commons Attribution Non-Commercial License 4.0 (CC BY-NC 4.0).
doi:10.3233/FAIA220703

Deep Transfer Learning Approach for

Obstructive Sleep Apnea Classification with
Photoplethysmography Signal
E SMILY JEYA JOTHIa,1J ANITHAb1, and D JUDE HEMANTH c
a
Assistant Professor, Department of Biomedical Instrumentation Engineering,
Avinashilingam Institute for Home Science and Higher Education for Women
b,c
Professor, Department of Electronics and Communication, Karunya Institute of
Technology and Sciences

[Link] health and quality of life are negatively impacted by apnea, an

increasingly prevalent sleep disorder. For monitoring and managing sleep apnea’s
side effects and consequences, accurate automatic algorithms for detecting sleep
apnea are crucial. In this paper, deep transfer learning methods are employed for
the detection of OSA events from Electrocardiograph (ECG) and
Photoplethysmography (PPG) signals. ResNet34 is a deep learning model based on
convolutional neural networks (CNNs). Transfer learning algorithms such as
AlexNet, VGG16, VGG19 and ResNet50 are implemented. In order to train the
ResNet34 model data augmentation, optimal learning rate finding, and fine-tuning
are used. To obtain generalizable models, a training set of data is divided into three
sets: a validation set for adjusting hyperparameters and improving generalizability,
and a test set for evaluating generalizability on unknown data. Deep transfer
learning models have the best accuracy, sensitivity, specificity, precision, and F1
score with 97.86±1.24%, 99.65%, 97.12%, 98.16% and 98.90% respectively. It
can assist sleep lab technicians in screening patients for OSA events continuously
through PPG and ECG signals.

[Link] classification, deep transfer learning, CNN, apneic events.

1. Introduction

A sleep disorder in which the upper airway partially or completely collapses is known
as obstructive sleep apnea (OSA). Research has demonstrated that OSA impairs
nocturnal sleep quality and causes fatigue during the day. OSA has also been linked to
cardiovascular and cerebrovascular disease, according to research [1]. Manual visual
annotation of PSG data is typically used to score sleep stages, including electrographic
measures of brain activity as well as eye and chin movements. An OSA patient can be
diagnosed using PSG, which measures apnea-hypopnea index (AHI) [5]. The severity
                      
   !"  #  $!"). An automated method of ECG analysis
could offer a potential solution to the problem of ECG abnormalities during overnight
sleep studies using Artificial Intelligence (AI) and machine learning (ML). Data
indicate that AI is capable of classifying sleep apnea more accurately than physicians in
categorizing ECG and PPG signals [2].

1
Dr. D Jude Hemanth
 E.S.J. Jothi et al. / Deep Transfer Learning Approach for Obstructive Sleep Apnea Classification 85

OSA diagnosis costs can be reduced through electrocardiogram (ECG)

measurements due to their simplicity. A disadvantage of ECG-based OSA events
detection is the complexity of the procedure, which requires qualified practitioners.
Detecting subtle changes in signal patterns can be made possible by computer
algorithms. Oxygen levels in the blood are reduced because OSA causes insufficient air
to enter the lungs. Sleep is frequently interrupted when patients lack oxygen in their
brains, which leads to frequent waking. Patients with cardiovascular disease are more
likely to suffer from OSA and engage in more cardiovascular risk factors. These
include hypertension, coronary heart disease, arrhythmia, heart failure, and stroke [5].
Observed Polysomnography (PSG) is the standard method of diagnosing sleep apnea.
Training technicians record several physiological signals and analyze them during
overnight hospitalizations in a sleep lab or sleep centre. In addition to being very
expensive, time-consuming, and limited in number, this approach involves a lot of
paperwork [6].

Saving resources and improving efficiency are two of the main benefits of
transfer learning. The use of neural networks to automatically learn features has been
proposed in several recent studies. Sparse auto-encoding was performed to encode the
unlabelled RR interval from a heartbeat signal. Fine tuning the classifier was
accomplished through discrimination in the learners' representations. A back
propagation network was optimized using the Newton method. A classifier was then
trained to extract the parameter of a Hidden Markov Model. Both the parameters and
probability outputs were sent into a Hidden Markov Model to estimate the probability.
In the end, different classifiers were combined using the decision fusion method [2].
The 1-min ECG signal was pre-processed with a Butterworth Band-pass filter of fourth
order with a frequency of 0.5 Hz to 15 Hz. A 1D CNN was used to identify normal and
apneic events from the pre-processed ECG signal. Tensor flow and Keras were used to
implement the proposed method. An ECG signal of one minute was used to extract
features using 10 identical feature layers. In order to feed the classifier, a two-
dimensional feature vector was converted into a one-dimensional feature vector. Based
on 1D feature vectors, four identical layers were designed to classify sleep apnea [3].
First, the ECG signals were pre-processed and segmented. Tests were conducted on a
variety of deep learning and machine learning methods. There was 92.27% specificity,
a sensitivity of 84.26 percent, and an accuracy of 88.13% for the hybrid deep model [4].

Using body sounds, an algorithm was developed to extract the heart rate.
Using the developed sleep monitor in conjunction with the PSG tests, ten subjects were
tested for validation. Apnea detection with the developed instrumentation was 92.8%
sensitive and 99.7% specific. The study did not report any significant signal losses [6].
An ECG signal was input to the CNN along with RR interval data and R peak data. A
60-second segment of non-overlapping segments was created for the ECG signal in the
pre-processing stage. Noise was reduced using an FIR band pass filter [8]. Using the
Hamilton algorithm, the R-peak in the ECG signal was identified, and uninterpretable
points were eliminated using a median filter. Per-segment and per-recording
classifications were accurate by 87.9% and 97.1%, respectively [9]. OSA events were
classified based on ECG scalograms and spectrograms. Continuous Wavelet Transform
and Short Time Fourier Transform were used to transform the ECG segments from the
time domain to the frequency domain.
86 E.S.J. Jothi et al. / Deep Transfer Learning Approach for Obstructive Sleep Apnea Classification

Predicting accuracy was compared between three different transfer learning models and
the proposed CNN model. ECG scalogram values were correctly predicted with
82.30% accuracy, sensitivity 82.27%, and specificity 82.95%, and ECG spectrogram
values were correctly predicted with 80.13% accuracy [15]. Based on 60 second
segment OSA classification, the network achieved 86.22% accuracy and 90%
sensitivity. Compared to ResNet, Chen et al. proposed an upgraded version of
pretrained network ResNet that improved accuracy by nearly 1% [16].

Accordingly, previous published work used pre-apnea segments of the brain

for feature extraction, and then used these features as inputs for traditional machine
learning algorithms for predicting apnea. We evaluated AlexNet, ResNet50, VGG16
and VGG19 networks in terms of their performance and comparative analysis in
identifying OSA segments based on PPG and ECG signals. Further, we demonstrated
that a hybrid deep learning architecture outperformed baseline experiments. PPG and
ECG recordings are found to be sufficient to predict OSA events, which makes the
proposed study novel compared to other prediction studies. This paper continues with a
discussion of the methodology in section 2, followed by a discussion and interpretation
of the results. The paper is concluded in section 3.

2. Methodology

2.1. Datasets

The deep transfer learning model on OSA detection is trained and evaluated
using the PhysioNet Apnea-ECG database from Philipps University, Marburg,
Germany. The Apnea-ECG dataset includes 70 single-channel ECG recordings, which
range from 401 to 587 minutes in length and are sampled at 100 Hz. There are 35
recordings in total, divided into two groups of 35 recordings each: available recordings
(a01 through a20, b01 through b05, and c01 through c10) and withheld recordings (x01
through x35). A one-minute segment of each recording was separated by non-
overlapping breaks. There are several files attached to each subject [10].

The Beth Israel Deaconess Medical Centre Patient PPG and Respiration Dataset is
publicly available in Physionet [11]. Data were collected from hospitalized critically ill
patients undergoing treatment at the Beth Israel Deaconess Medical Centre in Boston,
Massachusetts, USA. For each recording, two annotators manually annotated the
impedance respiratory signal. Each recording lasts 8 minutes and includes the
following information:

• Electrocardiograms, pulse oximetry, and impedance respiration signals are

examples of signs that represent bodily functions. Samples at 125 Hz are used.
• There are three physiological parameters: heart rate, respiration rate, and
blood oxygen saturation level. One sample is taken every second for each of
these parameters
• As well as variable variables like age and gender, fixed variables can be
recorded simultaneously.
• Manual annotations can be added to breaths
 E.S.J. Jothi et al. / Deep Transfer Learning Approach for Obstructive Sleep Apnea Classification 87

2.2. Transfer Learning

A CNN model can be built without creating a new one from scratch with
transfer learning. By utilizing and optimizing the knowledge of previously acquired
data, transfer learning provides training for CNN applications with small training
datasets [15]. The study investigates the efficiency of transfer learning on the entire
network using AlexNet, ResNet50, VGG16 and VGG19 models. The architecture of
these networks is shown in figure 1. In order to optimize weights and ensure
appropriate learning, learning rate is an influential hyperparameter. In addition, the
mini-batch size determines how many samples should be used during training in each
epoch. A single epoch was determined to be 0.0001, 64 epochs were determined, and a
maximum number of epochs was determined to be 30 epochs.

2.2.1. AlexNet

A total of eight layers are present in the AlexNet, each with a learnable
parameter. In each of the five layers except the output layer, they use ReLU activation
in combination with max pooling followed by three fully connected layers [4]. The
layers included are shown in table 1.

Table 1. Layer specification of AlexNet

Layer Filter Size Activation function
Input - -
Conv 1 11 x 11 ReLU
Max Pooling 1 3x3 -
Conv 2 5x5 ReLU
Max Pooling 2 3x3 -
Conv 3 3x3 ReLU
Conv 4 3x3 ReLU
Conv 5 3x3 ReLU
Max pooling 3 3x3 -
Dropout - -
Fully Connected 2

2.2.2. VGG16
With a dimension of 3 x 1, the VGG16 is composed of two convolutional
layers. In the next stage, there is a max pooling layer of 2 x 1, followed by two
convolutional layers with 128 kernels of size 3 x 1 stacked together. Finally, a
convolutional layer with 512 kernels of size 3 x 1 is included. Lastly, a pooling layer
with 2 x 1 is included in the fourth layer. Two layers consisting of 512 and 2 nodes
flatten the data before it is fed into two fully connected layers [4].

2.2.3. ResNet50
As part of its Artificial Intelligence research efforts, Google introduced the
ResNet50, a convolutional neural network with 50 layers and 7 million parameters. It is
trained with the help of the Physionet database. ResNet50 differs from other
architectures by having nine layers of inception. An organized convolutional layer
88 E.S.J. Jothi et al. / Deep Transfer Learning Approach for Obstructive Sleep Apnea Classification

consists of convolutional layers of 1x1, and 3x3 filters activated by ReLU functions,
each surrounded by a pooling layer. The output of every inception module concatenates
the acquired feature maps. In addition to reducing the number of parameters in the
network, ResNet50 also dramatically reduces its complexity. A global average pooling
is applied before the output layer, instead of a fully connected layer in traditional CNNs
[15, 17].

Figure [Link] of a) AlexNet, b) VGG16, c) VGG19, and d) ResNet50

2.2.4. VGG19
Table [Link] description of VGG19
Layers Kernel Size
2 Convolutional 64 3x1
Max-Pooling - 2x1
2 Convolutional 128 3x1
Max-Pooling - 2x1
4 Convolutional 256 3x1
Max-Pooling - 2x1
4 Convolutional 512 3x1
Max-Pooling - 2x1
4 Convolutional 512 3x1
Fully Connected 2
 E.S.J. Jothi et al. / Deep Transfer Learning Approach for Obstructive Sleep Apnea Classification 89

Table 2 shows the layer description of VGG19 [4].

2.3. Deep CNN ResNet34

CNN ResNet34 architecture [12] is used in this study to classify OSA events
via transfer learning. An existing trained model is taken and learned to solve a similar
classification problem instead of creating one from scratch. Based on the Physionet
database, the ResNet34 architecture is trained. Deep learning tasks are commonly
performed using this architecture. Convolutional neural networks are state-of-the-art
neural networks in computer vision that construct new models. Compared to other pre-
trained models such as Reception and VGG, the ResNet34 architecture converges faster,
and there is only one parameter to adjust. Comparing the architecture to other pre-
trained models like Inception and VGG, it is very simple to apply to different datasets.
The gradient approach cannot be used to update the weights of the first layers as the
number of layers increases. It transmits input data in a way that maintains gradients and
prevents information loss. When insufficient data is available to train a model, transfer
learning can provide significant benefits.
Convolutional neural networks consist of convolutional, pooling, and fully
connected layers. There are alternated convolutional and pooling layers in the
ResNet34 architecture. Convolutional base is a term used to describe this part of the
model. Rather than the ResNet34 model's dense layer, a new dense layer produces a
binary vector representing normal and abnormal classes. On top of the convolutional
base, the newly added dense layer is trained. In addition, a SoftMax activation function
is used in a final layer that returns logs of predictions instead of probabilities. There are
33 layers in the CNN convolutional base. These layers have a filter size of 3*3 and a
stride of 2 in the complete model. The hidden layers are activated using the rectified
linear unit (ReLU). Table 3 shows the number of parameters for each layer.

Table 3. Number of parameters for each layer

Layer Input size Output size Number of parameters
ResNet-34 3*128*128 64*64*64 9408
Dense 512 2 1026
SoftMax 2 2 0

2.3.1. Data Augmentation

It is generally the case that training a model with only a few data points leads
to over fitting during training. Using the validation set, the model generalizes but does
not preserve the details of the training set. This over fitting problem is mitigated by
employing the data augmentation technique during training. An approach to
augmenting existing training data is known as data augmentation. As a result, a novel
signal set for the training is created using data augmentation methods. Virtualizing data
through data augmentation techniques such as zooming, fusing, rotating, etc., is an easy
way to have more data.
90 E.S.J. Jothi et al. / Deep Transfer Learning Approach for Obstructive Sleep Apnea Classification

2.3.2. Optimal Learning Rate Finder

Model parameters are updated according to the gradient based on a hyper
parameter called learning rate. It takes a long time and produces tiny changes in the
model weights if the learning rate is set too low. The optimizer will overshoot the
minimum if the learning rate is too high, and it may even diverge if the learning rate is
too low. A model's performance is greatly affected by it [13].

2.3.3. Stochastic Gradient descent with restarts (SGDR)

The learning rate during training is moderately reduced by stochastic gradient
descent with restarts. With this approach, closer desired parameters can be identified,
so optimizations can be made with fewer changes to model parameters [14].

2.3.4. Fine-Tuning
Pre-trained models are fine-tuned by lightly adjusting their weights. ResNet34
is retrained with a new dense layer, as well as fine-tuned some parameters in the
convolutional base. A block diagram showing the proposed method for OSA
classification is shown in Figure 2. Raw data from PPG and ECG recordings is
retrieved from the Physionet database. As part of the data augmentation process, the
sample size is increased and the data is fed into the classifier. Having convolutional
base layers, ResNet34 is proposed as a deep transfer technique. Classifiers are tuned to
achieve the optimal learning rate. A classifier outputs two classes, OSA and Non-OSA,
according to the input signal.

Figure 2. Overview of the Proposed System

3. Discussions

The comparison between the proposed method and existing models can be found in
table 4. The pre-processing and the transfer learning techniques are implemented using
MATLAB R2020b environment. In order to optimize the parameters, a stochastic
gradient descent algorithm was used with a learning rate of 0.0001 in order to minimize
the cross-entropy loss function. After multiplying the learning rate and the bias of the
last fully connected layer, a constant is calculated. A set of measures, including
accuracy, sensitivity, specificity, precision, and F1 score, were calculated to evaluate
 E.S.J. Jothi et al. / Deep Transfer Learning Approach for Obstructive Sleep Apnea Classification 91

the accuracy, sensitivity, specificity, and precision of all models. The mathematical
expression can be formulated as in equation [1]-[5]
 
 = 100 (1)
 

 = 100 (2)


 = 100 (3)


 = 100 (4)


1  =  100 (5)
 ( )


A proposal for a prediction model for OSA detection and prediction is shown
in Table 4. This model achieves the highest accuracy of 97.86%, as well as sensitivity
of 99.65%, specificity of 97.12%, precision of 98.16% and F1 score of 98.90%. Other
pre-trained networks and existing methods are also summarized in the table. Rather
than improving OSA detection success, we sought to validate prediction performance
by comparing the proposed model with detection results on AlexNet, ResNet50,
VGG16, and VGG19. All models achieved the highest detection performance. For
prediction experiments, detection results were considered a reliable indicator. Figure 3
represents the graphical form of the performance of transfer learning and deep transfer
learning model.

Table 4. Comparison with the existing models

F1
Acc Sens Spec Precision
Reference Classifier Signal Score
(%) (%) (%) (%)
(%)
HMM,
DNN, Single lead
Li, K et al., [2] 85 88.9 - - -
Decision ECG
Fusion
Bahrami, M.,
&Forouzanfar, M. Deep CNN ECG 87.9 81.1 92.0 - -
[3]
Bahrami, M., Hybrid
&Forouzanfar, M. deep ECG 88.13 84.26 92.27 - -
[4] models
Gao, Q et al., [9] 1D CNN ECG 87.9 83.8 91.4 - -
Kalkbrenner, C et
- PPG - 92.8 99.7 - -
al., [6]
Radha, M et al., PPG and 76.36±7
LSTM - - - -
[7] ECG .57
Nasifoglu, H., ECG
CNN 82.30 83.22 82.27 82.95
+/;<=> scalograms
92 E.S.J. Jothi et al. / Deep Transfer Learning Approach for Obstructive Sleep Apnea Classification

ECG
80.13 81.99 77.87 - -
Spectrogram
CNN 97.86±0
99.65 97.12 98.16 98.90
ResNet34 .24
88.09±0
AlexNet 82.64 89.87 88.12 83.10
.89
PPG and 87.12±
Proposed model VGG16 82.06 90.50 90.78 83.13
ECG 0.36
86.75±
VGG19 81.44 90.07 87.15 82.52
0.38
88.13±
ResNet50 81.49 92.27 91.65 84.04
0.40

Figure 3. Performance of transfer learning and deep transfer learning model

There are some limitations to this study. The study excludes hypopnea, central sleep
apnea, and mixed sleep apnea as other types of sleep apnea. By having a larger set of
training and testing subjects, it would have been possible to evaluate models more
accurately. Because deep neural networks can automatically extract the most
important features, deep neural networks are superior to conventional machine learning.
In summary, the proposed approach has the following advantages:
x According to our knowledge, this is the first study to use deep
transfer learning to identify OSA events.
x A deep model that has proven its performance can be rapidly adapted
through deep transfer learning instead of creating a new model. As a
result, long validation processes are avoided as well as complicated
layer parameters.
x This study examines as many signals as possible.
 E.S.J. Jothi et al. / Deep Transfer Learning Approach for Obstructive Sleep Apnea Classification 93

4. Conclusion

An ECG and PPG signal classification method using deep transfer learning is presented
in this study. CNN-ResNet34 architecture is proposed for the automatic detection of
OSA events. It is comparable or better than existing deep learning methods as well as
traditional machine learning methods. In order to evaluate the model, 70 ECG records
and 53 PPG recordings are taken from the Physionet dataset. As an alternative to
manual scoring, this method can provide sleep technicians with an alternative way to
annotate SA events, and can promote clinical and household applications of SA
detection methods based on ECG and PPG signals.

References

[1] Gahungu, N., Shariar, A., Playford, D., Judkins, C., &Gabbay, E. Transfer learning artificial
intelligence for automated detection of atrial fibrillation in patients undergoing evaluation for suspected
obstructive sleep apnoea: a feasibility study. Sleep medicine, 85, 166-171. 2021
[2] Li, K., Pan, W., Li, Y., Jiang, Q., & Liu, G. A method to detect sleep apnea based on deep neural
network and hidden Markov model using single-lead ECG signal. Neurocomputing, 294, 2018
[3] Bahrami, M., &Forouzanfar, M.. Detection of Sleep Apnea from Single-Lead ECG: Comparison of
Deep Learning Algorithms. 2021 IEEE International Symposium on Medical Measurements and
Applications (MeMeA), 1-5. 2021
[4] Bahrami, M., &Forouzanfar, M. Sleep Apnea Detection From Single-Lead ECG: A Comprehensive
Analysis of Machine Learning and Deep Learning Algorithms. IEEE Transactions on Instrumentation
and Measurement, 71, 1-11. 2022
[5] Chang, H., Yeh, C., Lee, C., & Lin, C. A Sleep Apnea Detection System Based on a One-Dimensional
Deep Convolution Neural Network Model Using Single-Lead Electrocardiogram. Sensors (Basel,
Switzerland), 20. 2020
[6] Kalkbrenner, C., Eichenlaub, M.M., Rüdiger, S., Kropf-Sanchen, C., Rottbauer, W., &Brucher, R.
Apnea and heart rate detection from tracheal body sounds for the diagnosis of sleep-related breathing
disorders. Medical & Biological Engineering & Computing, 56, 671-681. 2017
[7] Radha, M., Fonseca, P., Moreau, A. et al. A deep transfer learning approach for wearable sleep stage
classification with photoplethysmography. npj Digit. Med. 4, 135 2021.
[8] Ramachandran, A., Pant, A., &Karuppiah, A Performance Analysis of Machine Learning Algorithms
for Sleep Apnea Detection Using ECG. Lecture Notes in Networks and Systems. 2021
[9] Gao, Q., Shang, L., & Zhang, Y. Automated Detection of Sleep Apnea Using Convolutional Neural
Network from a single-channel ECG signal. Proceedings of the 2020 2nd International Conference on
Robotics, Intelligent Control and Artificial Intelligence. 2020
[10] Penzel, T., Moody, G. B., Mark, R. G., Goldberge, A. L., Peter, J. H. 2000. The apnea-ECG data-base.
Computers in Cardiology 27, 255-258 (2000).
[11] Goldberger, A., Amaral, L., Glass, L., Hausdorff, J., Ivanov, P. C., Mark, R., J. E. Mietus, G. B. Moody,
C. K. Peng, and & Stanley, H. E. PhysioBank, PhysioToolkit, and PhysioNet: Components of a new
research resource for complex physiologic signals. Circulation [Online]. 101 (23), pp. e215–e220.
2000
[12] He, K., Zhang, X., Ren, S., & Sun, J. Deep residual learning for image recognition. In Proceedings of
the IEEE conference on computer vision and pattern recognition (pp. 770-778). 2016
[13] Smith, L. N. Cyclical learning rates for training neural networks. In Applications of Computer Vision
(WACV), 2017 IEEE Winter Conference on (pp. 464-472). IEEE, 2017
[14] Huang, G., Li, Y., Pleiss, G., Liu, Z., Hopcroft, J. E., & Weinberger, K. Q. Snapshot ensembles: Train 1,
get M for free. arXiv preprint arXiv:1704.00109, 2017
[15] ? < @ +/;< . Convolutional Neural Networks based OSA Event Prediction from ECG
Scalograms and Spectrograms. 2021
[16] Chen Q, Yue H, Pang X, Lei W, Zhao G, Liao E, Wang Y Mr-ResNeXt: A multi-resolution network
architecture for detection of obstructive sleep apnea. International Conference on Neural Computing for
Advanced Applications 420-432. 2020
[17] Rao, A.S., Nguyen, T., Palaniswami, M.S., & Ngo, T.D. Vision-based automated crack detection using
convolutional neural networks for condition assessment of infrastructure. Structural Health Monitoring,
20, 2124 - 2142. 2020