Computers in Biology and Medicine 113 (2019) 103387

Contents lists available at ScienceDirect

Computers in Biology and Medicine

journal homepage: www.elsevier.com/locate/compbiomed

Automated detection of diabetic subject using pre-trained 2D-CNN models T

with frequency spectrum images extracted from heart rate signals
⁎
Ozal Yildirima, , Muhammed Taloa, Betul Ayb, Ulas Baran Balogluc, Galip Aydinb,
U. Rajendra Acharyad,e,f
a
Department of Computer Engineering, Munzur University, Tunceli, Turkey
b
Department of Computer Engineering, Fırat University, Elazığ, Turkey
c
Department of Computer Science, University of Bristol, Bristol, United Kingdom
d
Department of Electronics and Computer Engineering, Ngee Ann Polytechnic, Singapore
e
Department of Biomedical Engineering, School of Science and Technology, Singapore School of Social Sciences, Singapore
f
School of Medicine, Faculty of Health and Medical Sciences, Taylor's University, 47500, Subang Jaya, Malaysia

A R T I C LE I N FO A B S T R A C T

Keywords: In this study, a deep-transfer learning approach is proposed for the automated diagnosis of diabetes mellitus
Diabetes mellitus (DM), using heart rate (HR) signals obtained from electrocardiogram (ECG) data. Recent progress in deep
Heart rate signals learning has contributed significantly to improvement in the quality of healthcare. In order for deep learning
Deep learning models to perform well, large datasets are required for training. However, a difficulty in the biomedical field is
Transfer learning
the lack of clinical data with expert annotation. A recent, commonly implemented technique to train deep
learning models using small datasets is to transfer the weighting, developed from a large dataset, to the current
model. This deep learning transfer strategy is generally employed for two-dimensional signals. Herein, the
weighting of models pre-trained using two-dimensional large image data was applied to one-dimensional HR
signals. The one-dimensional HR signals were then converted into frequency spectrum images, which were
utilized for application to well-known pre-trained models, specifically: AlexNet, VggNet, ResNet, and DenseNet.
The DenseNet pre-trained model yielded the highest classification average accuracy of 97.62%, and sensitivity of
100%, to detect DM subjects via HR signal recordings. In the future, we intend to further test this developed
model by utilizing additional data along with cloud-based storage to diagnose DM via heart signal analysis.

1. Introduction complications such as nephropathy (kidney), retinopathy (eye), cardi-

ovascular, and neuropathy (nerve) diseases [3–5]. While one of the
Diabetes Mellitus (DM) occurs when the blood glucose level is above most unnoticed complications is cardiovascular autonomic neuropathy
normal. Diabetes is a metabolic disorder that can happen at any age and (CAN), a least recognized and understood complication of diabetes is
cause serious complications. There are two major types of diabetes: diabetic autonomic neuropathy (DAN) [4]. CAN causes cardiac ab-
Type 1 and Type 2 [1]. Type 1 DM occurs when there is no insulin in the normalities; therefore monitoring of abnormalities using heart rate
body or minimal insulin production. Type 1 DM, which is usually seen variability (HRV) can help to detect at an early stage [5]. There are
in children and adolescents (< 30), causes coma, and there is a risk of various noninvasive techniques reported using fundus images to detect
death if not treated. Type 2 DM is a disease that usually occurs in diabetes and diabetic complications [6–13]. These studies provide re-
middle-aged or elderly patients (> 40). It occurs when the produced cognition using features from multiple regions [11,13] of fundus images
insulin is not consumed in the body, or sufficient insulin is not produced [6,7], tongue [8] and face [9,10,12].
by the body. Recently, many methods have been implemented successfully using
As reported by the International Diabetes Federation, 12% of global HRV signals to diagnose diabetes [14–20]. Acharya et al. [18] used
health spending is spent on diabetes, and one person dies every 6 s due discrete wavelet transform (DWT) features (of entropy and energy)
to diabetes [2]. In order to diagnose the disease, doctors use a blood obtained from the HRV signals and decision tree (DT) classifier to di-
glucose test. However, special care must be taken to prevent serious DM agnose diabetes. Their method yielded an accuracy of 92.02%,

⁎
Corresponding author.
E-mail address: [email protected] (O. Yildirim).

https://doi.org/10.1016/j.compbiomed.2019.103387
Received 4 June 2019; Received in revised form 8 August 2019; Accepted 8 August 2019
0010-4825/ © 2019 Elsevier Ltd. All rights reserved.
O. Yildirim, et al. Computers in Biology and Medicine 113 (2019) 103387

Fig. 1. Evolution of artificial intelligence (AI).

Fig. 2. A block representation of the proposed recognition system.

Fig. 3. Typical HR signal samples: a) Normal, and b) DM subjects.

sensitivity of 92.59%, and specificity of 91.46%. In another study [19], novel diabetes index approach for diagnosis of diabetic neuropathy
nonlinear features extracted from the HRV signals with the AdaBoost using features extracted from HRV signals. In Ref. [21], time, fre-
classifier obtained the highest average accuracy of 90%, sensitivity of quency, and nonlinear domain techniques were utilized to analyze
92.5%, and specificity of 88.7%. The same group [20] developed a normal and diabetic HR signals. They showed that nonlinear HRV

2
O. Yildirim, et al. Computers in Biology and Medicine 113 (2019) 103387

Fig. 4. Conversion of the original HR signals to frequency spectrum representative images: a) 1D HR signal, b) spectrogram of (a), and c) 2D grayscale image of (b).

analysis is more effective than time and frequency methods. Pachori [28–30], and it mimics the structure of the human brain. The emer-
et al. [22] classified diabetic and normal classes using features com- gence of new approaches and powerful computational resources to
puted from intrinsic mode functions (IMFs) obtained from the empirical compute and train the enormous amount of data have led to the rapid
mode decomposition (EMD) of RR-interval signals. Swapna et al. [23] growth in the development of deep neural networks. Fig. 1 shows the
used the higher order spectra (HOS) method on HR signals. Their evolution of artificial intelligence (AI). There are many applications of
technique obtained the maximum accuracy of 90.5% using Gaussian deep learning in biomedical image and signal processing studies
mixture model (GMM). Using linear regression, Nolan et al. [24] per- [31–37]. Pratt et al. [38] have used a convolutional neural network
formed a gender-based relationship analysis between HRV measures (CNN) to classify diabetic retinopathy (DR) stages. Their network has
and the duration of type 2 diabetes. They reported gender-based dis- reached a classification accuracy of 75% using 5 K validation images.
tinctions among vagal-heart rate modulation, duration of diabetes, and In this study, we used deep learning CNN models for the detection of
total R-R variability in the HRV signals. Trunkvalterova et al. [25] diabetic subjects using HR signals. We have used the most popular deep
employed the multiscale entropy (MSE) analysis to detect subtle ab- learning pre-trained models: AlexNet, VggNet, DenseNet, and ResNet,
normalities in young type 1 diabetes patients’ cardiovascular system. trained using large image datasets, to achieve a higher detection per-
Seyd et al. [26] applied frequency and time approaches to discern formance. We transformed the HR signals into spectrogram images for
normals from DM patients by analyzing HR signals. For time and fre- pre-trained models. In this way, we achieved a significant improvement
quency domain analysis, they have used the ECG signals of 16 DM in classification performance. To the best of our knowledge, the pro-
patients and 16 normal subjects. Mercaldo et al. [27] utilized different posed study is the first work to apply the 2-dimensional deep transfer
machine learning methods to differentiate diabetes affected patients learning approach using 1-dimensional HR signal data. The overall
from controls. Using the Hoeffding tree algorithm, they obtained a contribution of this study is summarized as follows:
precision value of 77%.
The classical machine learning methods used to diagnose diabetes - Provided an effective classification of DM subject with a complete
have difficulties. Feature extraction is one of the most important steps end-to-end structure without requiring any hand-crafted feature
in traditional machine learning systems. The performance of the ma- extraction techniques.
chine learning system depends on feature extraction. Extraction of the - A deep learning-based approach has been developed using HR sig-
best performing features is done by the trial and error method, which is nals.
time-consuming. Deep learning performs automatic feature learning - Spectrogram images enabled pre-trained deep learning models to be

3
O. Yildirim, et al. Computers in Biology and Medicine 113 (2019) 103387

Fig. 5. Graphical representation of spectrogram images and respective HR signals: a) Normal subject b) DM subject.

Fig. 6. Graphical representation of the test and training data sets for the 1D-CNN model.

trained on a small dataset (71 normal and 71 DM). more visual representations. For this purpose, 1-dimensional signal data
- Using deep transfer learning, the difficulties in the stages of model is converted to 2-dimensional gray images by the Short-Time Fourier
training and design is eliminated. Transform (STFT) method. The images having visual representations of
the frequency spectra are used to train and test various popular pre-
trained models. The block representation of the proposed method is
2. Material and methods
shown in Fig. 2.

In this study, a deep learning framework is proposed for the de-

tection of DM using HR signals. In order to benefit from the perfor- 2.1. HR dataset
mance of pre-trained deep learning models which have been trained on
the ImageNet database, HR signals are transformed into images having The ECG signals used in the study were obtained from 15 DM

4
O. Yildirim, et al. Computers in Biology and Medicine 113 (2019) 103387

Table 1 EEG signals [44]. Salem et al. [45] have used spectrogram images of
The layers and layer parameters of the 1D-CNN model designed for this work. ECG signals for classification based on deep transfer learning. In our
No Layer Name Filter Size Kernel Size Other Parameters Output Size present study, we have converted HR signals to spectrogram images for
use with pre-trained 2D CNN models. In Fig. 4, the flow of the con-
0 Input Layer – – – 1000 × 1 version of the original HR signals to spectrogram images is provided.
1 Conv-1D 64 5 Strides = 3 332 × 64
HR signals with 1000 samples after passing through STFT are con-
2 Conv-1D 64 3 Strides = 1 330 × 64
3 Max-Pooling Strides = 2, Pool 165 × 64
verted into spectrograms, which represent the frequency spectra. The
size = 2 spectrogram indicates changes in the frequency spectra of the original
4 Dropout – – Rate = 0.1 165 × 64 signal, which can be observed over time. The grayscale images (Fig. 4
5 Conv-1D 128 3 Strides = 1 163 × 128 (c)) of STFT images of size 437 × 501 pixels is presented. Fig. 5 shows
6 Conv-1D 128 5 Strides = 1 159 × 128
the original HR signals of two normal and two DM subjects and the
7 Max-Pooling – – Strides = 2, Pool 79 × 128
size = 2 corresponding spectrogram images.
8 Conv-1D 256 2 Strides = 1 78 × 256
9 Conv-1D 256 3 Strides = 1 76 × 256 2.3. Deep networks
10 Max-Pooling – – Strides = 2, Pool 38 × 256
size = 2
11 Conv-1D 64 3 Strides = 1 36 × 64
Deep learning is of great importance in many areas as well as in the
12 Conv-1D 64 5 Strides = 1 32 × 64 medical field. Deep learning models, especially CNNs, have been suc-
13 Max-Pooling – – Strides = 2, Pool 16 × 64 cessfully used in medical applications such as detection [46] and clas-
size = 2 sification [47]. Krizhevsky et al. [48] ranked the first by successfully
14 Flatten – – – 1024
classifying images in the ImageNet large-scale visual recognition chal-
15 Dense (Relu) – – Hidden units = 64 64
16 Dense – – Hidden units = 2 2 lenge (ILSVRC) with the developed deep learning-based CNN model
(Softmax) (AlexNet) in 2012. In 2014, the VGGNet [49] CNN model demonstrated
a better classification performance than AlexNet in ILSVRC. In the
*Hyper Parameters: Optimizer = Adam, Batch size = 8, Learning rate = 0.001 following year, the ResNet model developed by He et al. [50], which
won the first place award for classification, detection, and segmentation
patients (7 female and 8 male) and 15 normal (5 female and 10 male) tasks at ILSVRC 2015. The ResNet model was developed with the skip
subjects for a 1 h duration [18]. In this study, diabetes patients with the connection technique, which is deeper than the previous models
disease for 5–15 years and age ranging from 50 to 70 years were con- (AlexNet and VggNet). The skip connections, also known as residual
sidered. The subjects in the normal group were between 40 and 60 connections, attach every residual block to the next blocks. This tech-
years old. The ECG signals were recorded at Kasturba Medical Hospital nique enables information flow throughout the network. Therefore, it
(KMH), Manipal, India. The pre-processing steps performed on the allows training of the CNN even with 1000 layers. Huang et al. [51]
signals were as follows: introduced densely connected convolutional networks (DenseNet) using
a similar shortcut technique which connects every layer of the network
• Signals were recorded at 500 Hz sampling rate using the BIOPACTM to the following layers. DenseNet architecture generally showed high
system. performance in classification problems such as ResNet architecture.
• 15 Hz cut-off frequency low-pass filter for noise removal and 0.3 Hz However, DenseNet architecture used fewer parameters and required
cut-off frequency high-pass filter to remove baseline wander were fewer computations than ResNet in training the model. Besides, one of
used. the advantages of DenseNet model is that it exhibits better classification
• A band reject filter (50 Hz central frequency) was applied to clear performance using small datasets [51].
away the power-line interfacing noise. The transfer learning method plays a vital role in finding a solution
• The Pan-Tompkins algorithm [39] was used for RR point detection. for classification problems. A CNN usually requires a large dataset to
train with a large number of labeled data, and it requires higher com-
A total of 142 data files with each file containing 1000 samples were putational power. One of the main problems in medical data analysis is
used. Hence, we have used 71 DM and 71 normal data files in this work. the limited number of annotated data. The process of labeling data by
Fig. 3 shows the sample HR signals belonging to normal and DM sub- experts is expensive and time-consuming. Therefore, instead of training
jects. the CNN model from scratch, we used the weights of the pre-trained
models which already learned the distinguishing representations from a
2.2. Spectrogram images different but similar task. In other words, in the transfer learning
technique, the weights of pre-trained models are transferred to the
The Fourier transform (FT) is a crucial methodology which is not present model. In this way, small datasets are trained with a low
useful when the signal is nonstationary, and its spectral content is computational cost.
changing [40]. The HR signals represent nonstationary behavior by
having altered frequencies and amplitudes over time. This change 3. Experimental results
cannot be captured by FT based analysis. Therefore, we preferred to use
a method which could capture these changes. The short-time Fourier In this study, HR signals are used for the detection of diabetes pa-
transform (STFT) is a general-purpose FT based function, which can tients, and these signals are classified using deep learning based ap-
help the deep learning structure to extract hidden features from the proaches. We have performed this in two steps. First, a CNN model is
spectrogram images effectively. STFT has been used in the past with designed for raw HR signals, and the performance of this model is in-
CNN, and Long Short-Term Memory (LSTM) architectures for speech vestigated. In the second step, we have tried to increase detection
recognition [41] and motor imagery brain-computer interface re- performance by examining the cases where the developed CNN model is
cognition [42] tasks. The speech signal is a time-varying signal, and inadequate. Therefore, the input signals are converted to image data-
STFT can create a 2-dimensional representation of this signal. From the sets. Then, the obtained images are classified with pre-trained models.
segmented audio data, STFT is used to generate binary images of speech The training and testing of the proposed deep models are carried out
and music [43]. Similarly, STFT can be used to generate a 2-dimen- on a Linux server with the Ubuntu 16.04 operating system using 11 GB
sional representation of HR signals. In another signal classification of memory, including an NVIDIA GeForce GTX 1080 TI graphics card.
study, the STFT and CNN combination has been successfully applied to The results are evaluated using the k-fold cross-validation (CV)

5
O. Yildirim, et al. Computers in Biology and Medicine 113 (2019) 103387

Fig. 7. The performance graphs of 1D CNN model obtained during 100 epochs: a) Accuracy, and b) Loss.

Table 2 model is composed of convolution, pooling, and dense layers, and in-
1D CNN model's performance values for 5-fold test data. cludes a dropout layer to prevent overfitting. Table 1 presents the layer
Fold Number Accuracy (%) Sensitivity (%) Precision (%) F1-Score
parameters of the proposed model. The model parameters are adjusted
by the brute-force technique, similar to our previous deep learning
Fold-1 86.21 100 69.23 81.82 works done using ECG and EEG signal data [31,33,46,47]. The Keras
Fold-2 79.31 85.71 75.00 80.00 deep learning library is used to construct, train, and test the model.
Fold-3 86.21 84.21 94.12 88.89
Fold-4 89.66 100 80.00 88.89
In this work, 80% of the HR signals were used for training, and the
Fold-5 89.66 92.86 86.67 89.66 remaining 20% for validation. The performance evaluations were made
Mean ± Std. 86.21 ± 4.22 92.55 ± 7.54 81.00 ± 9.74 85.85 ± 4.56 using a 5-fold CV strategy. The training and validation accuracy and
loss graphs of the model for a fold for 100 epochs are given in Fig. 7.
The CNN model completed the training process without having any
strategy. Thus, one of the folds is used as a validation set, and the re- overfitting problem on one dimensional HR signals. After training, the
maining folds are used as the training set. The k value for HR signals is training accuracy reached 86.73%, while the validation accuracy re-
set to 5. Therefore, 20% of the data is reserved for testing, while the mained at 86.21%. Table 2 presents the performance values of the CNN
remaining 80% is used in the training phase. Fig. 6 shows the block network using a 5-fold CV strategy. The CNN network can reach
representation of the training and testing dataset presented to the 86.21% ± 4.22 average accuracy with 5-fold CV. In the end, we con-
proposed models. cluded that the 1-dimensional CNN model did not reach the desired
level of success.
3.1. 1D CNN model
3.2. Pre-trained 2D CNN models
For the classification of HR signals, a 16-layer CNN model is con-
structed. The 1000 × 1 HR signals are input to this model. The CNN The most critical stage of the experimental studies is converting HR

6
O. Yildirim, et al. Computers in Biology and Medicine 113 (2019) 103387

Fig. 8. Sample images obtained after converting HR signals to frequency spectrum images: a) Normal subject, and b) DM subject.

trained only the last layers (fully-connected) of pre-trained models.

Fig. 9 shows the training performance of pre-trained models using the
same-fold data.
The proposed pre-trained models are trained for 40 epochs. When
the performances of the pre-trained models are evaluated, we witnessed
that the DenseNet pre-trained model obtained the best classification
accuracy among others. While the performance of AlexNet and VggNet
are similar, the lowest performance is achieved with ResNet. The sen-
sitivity, specificity, precision, F1-score, and accuracy measurements are
used to evaluate the performance of the pre-trained models. The per-
formance values of the proposed pre-trained models for the 5-fold cross-
validation (CV) strategy are given in Table 3.
The classification performances of the AlexNet and ResNet-34
models are interestingly similar. The average accuracy values of both
networks showed the lowest performance of 95.7% ± 1.0 The highest
Fig. 9. Graphs of accuracies obtained for different epochs with various pre-
average accuracy of 97.6% ± 2.3 is obtained using DenseNet-161. In
trained models: AlexNet (blue line), DenseNet (orange line), ResNet (green line)
addition, the sensitivity value of DenseNet-161 is increased to as high as
and VggNet (purple line). (For interpretation of the references to colour in this
100% with the 5-fold CV strategy. Fig. 10 shows the accuracy and loss
figure legend, the reader is referred to the Web version of this article.)
graphs of the DenseNet-161 model during 40 epoch periods for a fold.
It can be observed from the accuracy graph (Fig. 10(a)) that Den-
signals into images that can be processed with pre-trained 2D-CNN seNet-161 model reached the highest classification accuracy of 100% at
models to improve classification performance. Popular pre-trained the end of 40 epochs for the test images. When the loss graphs are
models such as AlexNet, VggNet, DenseNet, and ResNet are trained and analyzed (Fig. 10(b)), both training and validation loss values de-
tested on these image data. For this purpose, the STFT method is used creased during the training of the model. It may be noted that, when the
on 1-dimensional HR signals to obtain 2-D images, which indicate the training of the DenseNet-161 model is completed, the validation loss is
visual representations of frequency spectra. Fig. 8 shows a few images still less than the training loss (underfitting). The main reason for this is
that represent frequency spectra corresponding to normal and DM that the DenseNet-161 has correctly classified each image in the vali-
subjects. dation set; therefore, the classification accuracy of the model has
It is difficult to discriminate the two classes (normal/DM) using yielded 100%, and the validation loss reached the minimum value. It
spectrogram images visually. Deep learning methods can perform the can be observed from this result that the model's predictions on the
recognition process with high performance by extracting abstract fea- images in the validation set, which it has not seen before during
tures from these images. The performance of the networks can be sig- training, were very successful. This outcome assured that the training of
nificantly increased, especially with the use of pre-trained models the model was at the desired level. Fig. 11 shows the confusion matrix
which are trained on large data. In our study, we used AlexNet, acquired with the 5-fold cross validation (CV) strategy using the Den-
DenseNet, ResNet, and VggNet pre-trained models, which performed seNet-161 model with diabetes test data.
well for image classification tasks in the field of deep learning. We have The model misclassified only five images in the DM class, and

Table 3
The performance values obtained by various pre-trained models with 5-fold test data (mean ± standard deviation).
Models Sensitivity (%) Specificity (%) Precision (%) F1-Score (%) Accuracy (%)

AlexNet 93.66 ± 3.6 96.60 ± 0.1 92.13 ± 0.7 92.87 ± 1.9 95.72 ± 1.0
VggNet (16) 95.26 ± 4.3 97.29 ± 2.8 94.13 ± 5.7 94.51 ± 2.0 96.67 ± 1.3
ResNet (34) 93.80 ± 3.5 96.59 ± 0.1 92.13 ± 0.7 92.93 ± 1.6 95.72 ± 1.0
DenseNet (161) 100 96.72 ± 3.3 92.33 ± 7.6 95.88 ± 4.1 97.62 ± 2.3

7
O. Yildirim, et al. Computers in Biology and Medicine 113 (2019) 103387

Fig. 10. The performance graphs of DenseNet-161 model on the diabetes test dataset: a) testing accuracy b) training and validation loss.

pre-trained models.
Our findings show that the results obtained using pre-trained
models have yielded better classification accuracy using 2-dimensional
data than the proposed 16-layer deep learning model with 1-dimen-
sional data. The AlexNet, VggNet, ResNet, and DenseNet pre-trained
CNN models have been trained with more than a million images be-
longing to 1000 different categories. These CNN models have already
learned valuable representations from various images. We have used
this gained knowledge by implementing the transfer learning technique
for 2-dimensional data. For 1-dimensional data, the proposed deep
learning model only learned the representations from the given data.
Therefore, the classification accuracy of the constructed model is lower
than for the pre-trained CNN models. In the future, we hope that such
transfer learning techniques can also be employed for 1-dimensional
data to obtain high performance.
Fig. 11. Overall confusion matrix obtained with 5-fold for the DenseNet-161
model.
4. Discussion
correctly classified all normal class images in all folds. The 5-fold
There are a few remarkable state-of-the-art studies performed for
average classification accuracy value is 97.6% ± 2.3. The performance
the automated detection of DM subjects using HR signals. In these
of the DenseNet-161 model on each fold data is given in Table 4.
studies, the discrete wavelet transform (DWT) [18], empirical mode
The developed 16-layer CNN model has been trained from scratch.
decomposition (EMD) [22], higher order spectra (HOS) [23], non-linear
Initially, the weights of the proposed network were set randomly and
analysis [21], statistical methods [20] have been employed for feature
then adjusted through backpropagation. The whole network is trained
extraction. For automated detection of DM, obtained features are input
until optimal performance is reached. However, for our 2-dimensional
to the shallow-structured classifiers, such as decision tree (DT) [18],
data, we have used pre-trained models (AlexNet, VggNet, ResNet, and
AdaBoost [19,20], Gaussian mixture model (GMM) [23], and support
DenseNet) to detect a diabetic subject using frequency spectrum
vector machines (SVM) [49]. A few other signal processing methods,
images. These pre-trained models have already been trained to perform
coupled with machine learning models, have been used to detect DM
other classification tasks. We have transferred the weights of the pro-
automatically [21,22]. The accuracies obtained using these studies
posed pre-trained models. During the training of pre-trained CNN
have a range between 90% and 92%. Our present study employed a
models, we have only trained the fully connected layers of the pre-
deep learning-based approach for DM detection. With this completely
trained models. Therefore, the deep learning model used for 1-dimen-
end-to-end structure, the signals are classified without requiring any
sional data trained from scratch requires more epochs (100) than the
handcrafted feature extraction. Table 5 shows a comparison of

Table 4
The performances of DenseNet-161 model for each of 5-fold DM classification.
Fold Number Sensitivity (%) Specificity (%) Precision (%) F1-Score (%) Accuracy (%)

Fold-1 100 100 100 100 100

Fold-2 100 96.77 91.67 95.65 97.62
Fold-3 100 93.10 86.67 92.86 95.24
Fold-4 100 93.75 83.33 90.91 95.24
Fold-5 100 100 100 100 100

8
O. Yildirim, et al. Computers in Biology and Medicine 113 (2019) 103387

Table 5
Comparison of performances for automated detection of diabetes using the same HRV signal database.
Study Methods Results/Findings

Acharya et al. [18] DWT & Decision Tree Sen = 92.59%, Acc = 92.02%
Acharya et al. [19] Nine nonlinear measures & AdaBoost Sen = 92.50%, Acc = 90.00%
Acharya et al. [20] Diabetic integrated index & AdaBoost Sen = 87.50%, Acc = 86.00%
Faust et al. [21] Six non-linear features & student t-test Non-linear analysis is more effective than frequency and time domain analysis
Pachori et al. [22] EMD & Kruskal–Wallis statistical test Significant difference between diabetic and normal classes (p < 0.05)
Swapna et al. [23] HOS & GMM Sen = 85.70%, Acc = 90.50%
Jian and Lim [52] HOS & SVM Sen = 70.97%, Acc = 79.93%
This study End-to-end 1D HR signals & 1D-CNN model Sen = 92.55% ± 7.5,
Acc = 86.21% ± 4.2
This study End-to-end 2D frequency spectrum images & 2D-CNN DenseNet161 Sen = 100%,
Acc = 97.62% ± 2.3

performance for automated detection of diabetes using the same HRV www.diabetesatlas.org/resources/2017-atlas.html.
signal database. The proposed study achieved a 97.62% ± 2.3 average [3] W.T. Cade, Diabetes-related microvascular and macrovascular diseases in the
physical therapy setting, Phys. Ther. 88 (11) (2008) 1322–1335.
accuracy and a sensitivity of 100%, which outperformed existing state- [4] A.I. Vinik, R.E. Maser, B.D. Mitchell, R. Freeman, Diabetic autonomic neuropathy,
of-the-art studies in the literature. Diabetes Care 26 (5) (2003) 1553–1579.
The advantage of the present study is the transfer of weights from [5] A.I. Vinik, D. Ziegler, Diabetic cardiovascular autonomic neuropathy, Circulation
115 (3) (2007) 387–397.
popular models, which trained on two-dimensional large image data, to [6] M. Niemeijer, et al., Retinopathy online challenge: automatic detection of micro-
a small number of signal data. Thus, the constraint in the construction aneurysms in digital color fundus photographs, IEEE Trans. Med. Imaging 29 (1)
and training phase of deep models is eliminated. Also, such a developed (2009) 185–195.
[7] U.R. Acharya, C.M. Lim, E.Y.K. Ng, C. Chee, T. Tamura, Computer-based detection
model (in our present work) has yielded higher classification perfor-
of diabetes retinopathy stages using digital fundus images, Proc. Inst. Mech. Eng. H
mance. In the future, we can use this model to detect diabetes at an J. Eng. Med. 223 (5) (2009) 545–553.
early stage, and it also can be used to detect other diseases from bio- [8] B. Zhang, B.V. Kumar, D. Zhang, Detecting diabetes mellitus and nonproliferative
diabetic retinopathy using tongue color, texture, and geometry features, IEEE Trans.
medical signals. The main disadvantage of this work is that we have
Biomed. Eng. 61 (2) (2013) 491–501.
used a small database. Deep learning models consist of many layers [9] B. Zhang, D. Zhang, Noninvasive diabetes mellitus detection using facial block color
with millions of parameters in these layers. The models process the data with a sparse representation classifier, IEEE Trans. Biomed. Eng. 61 (4) (2013)
iteratively to obtain optimal parameters during the training phase. 1027–1033.
[10] T. Shu, B. Zhang, Y.Y. Tang, An extensive analysis of various texture feature ex-
Repeated processing of data results in various problems with small data tractors to detect Diabetes Mellitus using facial specific regions, Comput. Biol. Med.
because the model memorizes the training data and fails to recognize 83 (2017) 69–83.
test data, which it has not seen (overfitting problem). This overfitting [11] J. Li, D. Zhang, Y. Li, J. Wu, B. Zhang, Joint similar and specific learning for dia-
betes mellitus and impaired glucose regulation detection, Inf. Sci. 384 (2017)
problem can be addressed using a transfer learning technique by em- 191–204.
ploying the models that have been previously trained on large datasets. [12] T. Shu, B. Zhang, Y.Y. Tang, An improved noninvasive method to detect Diabetes
With the proposed method, it is possible to optimize the existing Mellitus using the Probabilistic Collaborative Representation based Classifier, Inf.
Sci. 467 (2018) 477–488.
parameters (learned features) instead of setting all model parameters to [13] J. Li, B. Zhang, G. Lu, J. You, D. Zhang, Body surface feature-based multi-modal
zero. learning for diabetes mellitus detection, Inf. Sci. 472 (2019) 1–14.
[14] U.R. Acharya, H. Fujita, V.K. Sudarshan, V.S. Sree, L.W.J. Eugene, D.N. Ghista,
R. San Tan, An integrated index for detection of sudden cardiac death using discrete
5. Conclusion
wavelet transform and nonlinear features, Knowl. Based Syst. 83 (2015) 149–158.
[15] Y. Cheng, F. Wang, P. Zhang, J. Hu, Risk prediction with electronic health records: a
In this study, we proposed a deep transfer learning based approach deep learning approach, Proc. 2016 SIAM International Conference on Data Mining,
2016, pp. 432–440.
using spectrogram images obtained from HR signals to detect diabetes
[16] M.R.K. Mookia, U.R. Acharya, C.K. Chua, C.M. Lim, E.Y.K. Ng, A. Laude, Computer-
patients automatically. AlexNet, VggNet, ResNet, and DenseNet CNN aided diagnosis of diabetic retinopathy: a review, Comput. Biol. Med. 43 (12)
pre-trained models trained on 2D image data were used for the eva- (2013) 2136–2155.
luation of one-dimensional HR signals. A total of 142 segments (71 [17] M.R.K. Mookiah, U.R. Acharya, R.J. Martis, C.K. Chua, C.M. Lim, E.Y.K. Ng,
A. Laude, Evolutionary algorithm based classifier parameter tuning for automatic
normal and 71 DM) obtained from 30 subjects (15 normal and 15 DM) diabetic retinopathy grading: a hybrid feature extraction approach, Knowl. Based
were used in this study. The DenseNet-161 CNN model achieved a Syst. 39 (2013) 9–22.
97.62% ± 2.3 accuracy and a 100% sensitivity performance with the [18] U.R. Acharya, K.S. Vidya, D.N. Ghista, W.J.E. Lim, F. Molinari,
M. Sankaranarayanan, Computer-aided diagnosis of diabetic subjects by heart rate
5-fold CV strategy. Hence, diabetes patients can be discerned accurately variability signals using discrete wavelet transform method, Knowl. Based Syst. 81
using the pre-trained 2D-CNN models with frequency spectrum images (2015) 56–64.
extracted from heart rate signals. In the future, we intend to improve [19] U.R. Acharya, O. Faust, N.A. Kadri, J.S. Suri, W. Yu, Automated identification of
normal and diabetes heart rate signals using nonlinear measures, Comput. Biol.
model accuracy by training it with more data and also focus on the Med. 43 (10) (2013) 1523–1529 2013.
early detection of diabetes. We also plan to explore the possibility of [20] U.R. Acharya, O. Faust, S.V. Sree, D.N. Ghista, S. Dua, P. Joseph, V.I.T. Ahamed,
using more HR signals to evaluate the performance of our developed N. Janarthanan, T. Tamura, An integrated diabetic index using heart rate variability
signal features for diagnosis of diabetes, Comput. Methods Biomech. Biomed. Eng.
model. 16 (2) (2013) 222–234.
[21] O. Faust, U.R. Acharya, F. Molinari, S. Chattopadhyay, T. Tamura, Linear and non-
Conflicts of interest linear analysis of cardiac health in diabetic subjects, Biomed. Signal Process.
Control 7 (3) (2012) 295–302.
[22] R.B. Pachori, P. Avinash, K. Shashank, R. Sharma, U.R. Acharya, Application of
All authors declare that there is no conflict of interest in this work. empirical mode decomposition for analysis of normal and diabetic RR-interval
signals, Expert Syst. Appl. 42 (9) (2015) 4567–4581.
References [23] G. Swapna, U.R. Acharya, S. VinithaSree, J.S. Suri, Automated detection of diabetes
using higher order spectral features extracted from heart rate signals, Intell. Data
Anal. 17 (2) (2013) 309–326.
[1] I. Benjamin, R.C. Griggs, E.J. Wing, J.G. Fitz, Andreoli and Carpenter's Cecil [24] R.P. Nolan, S.M. Barry-Bianchi, A.E. Mechetiuc, M.H. Chen, Sex-based differences in
Essentials of Medicine E-Book, Elsevier Health Sciences, 2015. the association between duration of type 2 diabetes and heart rate variability,
[2] International Diabetes Federation, IDF Diabetes Atlas, eighth ed., (2017) http:// Diabetes Vasc. Dis. Res. 6 (4) (2009) 276–282.

9
O. Yildirim, et al. Computers in Biology and Medicine 113 (2019) 103387

[25] Z. Trunkvalterova, M. Javorka, I. Tonhajzerova, J. Javorkova, Z. Lazarova, [40] S. Nisar, O.U. Khan, M. Tariq, An efficient adaptive window size selection method
K. Javorka, M. Baumert, Reduced short-term complexity of heart rate and blood for improving spectrogram visualization, Comput. Intell. Neurosci. 2016 (2016).
pressure dynamics in patients with diabetes mellitus type 1: multiscale entropy [41] W. Lim, D. Jang, T. Lee, Speech emotion recognition using convolutional and re-
analysis, Physiol. Meas. 29 (7) (2008) 817. current neural networks, 2016 Asia-Pacific Signal and Information Processing
[26] P.A. Seyd, V.T. Ahamed, J. Jacob, P. Joseph, Time and frequency domain analysis of Association Annual Summit and Conference (APSIPA), 2016, pp. 1–4.
heart rate variability and their correlations in diabetes mellitus, Int. J. Biol. Life Sci. [42] Z. Wang, L. Cao, Z. Zhang, X. Gong, Y. Sun, H. Wang, Short time Fourier trans-
4 (1) (2008) 24–27. formation and deep neural networks for motor imagery brain computer interface
[27] F. Mercaldo, V. Nardone, A. Santone, Diabetes mellitus affected patients classifi- recognition, Concurrency Comput. Pract. Ex. 30 (23) (2018) e4413.
cation and diagnosis through machine learning techniques, Procedia comput. sci. [43] P. Neammalai, S. Phimoltares, C. Lursinsap, Speech and music classification using
112 (2017) 2519–2528. hybrid form of spectrogram and fourier transformation, Signal and Information
[28] Y. LeCun, Y. Bengio, G. Hinton, Deep learning, Nature 521 (7553) (2015) 436. Processing Association Annual Summit and Conference(APSIPA), 2014, pp. 1–6.
[29] I. Goodfellow, Y. Bengio, A. Courville, Deep Learning, MIT press, 2016. Asia-Pacific 2014.
[30] M. Coşkun, Ö. Yildirim, A. Uçar, Y. Demir, An overview of popular deep learning [44] Y. Lu, H. Jiang, W. Liu, Classification of eeg signal by stft-cnn framework: identi-
methods, Eur J Tech 7 (2) (2017) 165–176. fication of right-/left-hand motor imagination in bci systems, The 7th International
[31] O. Yildirim, U.B. Baloglu, U.R. Acharya, A deep learning model for automated sleep Conference on Computer Engineering and Networks, vol 299, 2017, p. 1.
stages classification using psg signals, Int. J. Environ. Res. Public Health 16 (4) [45] M. Salem, S. Taheri, J.S. Yuan, ECG arrhythmia classification using transfer learning
(2019) 599. from 2-dimensional deep CNN features, 2018 IEEE Biomedical Circuits and Systems
[32] S.K. Pandey, R.R. Janghel, Recent deep learning techniques, challenges and its Conference (BioCAS), 2018, pp. 1–4.
applications for medical healthcare system: a review, Neural Process. Lett. (2019) [46] Ö. Yıldırım, P. Pławiak, R.S. Tan, U.R. Acharya, Arrhythmia detection using deep
1–29. convolutional neural network with long duration ECG signals, Comput. Biol. Med.
[33] U.B. Baloglu, M. Talo, O. Yildirim, R. San Tan, U.R. Acharya, Classification of 102 (2018) 411–420.
myocardial infarction with multi-lead ECG signals and deep CNN, Pattern Recognit. [47] M. Talo, U.B. Baloglu, Ö. Yıldırım, U.R. Acharya, Application of deep transfer
Lett. 122 (2019) 23–30. learning for automated brain abnormality classification using MR images, Cogn.
[34] S.L. Oh, E.Y. Ng, R. San Tan, U.R. Acharya, Automated beat-wise arrhythmia di- Syst. Res. 54 (2019) 176–188.
agnosis using modified U-net on extended electrocardiographic recordings with [48] A. Krizhevsky, I. Sutskever, G.E. Hinton, Imagenet classification with deep con-
heterogeneous arrhythmia types, Comput. Biol. Med. 105 (2019) 92–101. volutional neural networks, Advances in Neural Information Processing Systems,
[35] O. Faust, Y. Hagiwara, T.J. Hong, O.S. Lih, U.R. Acharya, Deep learning for 2012, pp. 1097–1105.
healthcare applications based on physiological signals: a review, Comput. Methods [49] Simonyan K, Zisserman A. (2014) Very Deep Convolutional Networks for Large-
Progr. Biomed. 161 (2018) 1–13. Scale Image Recognition. arXiv preprint 2014; arXiv:1409.1556.
[36] N. Michielli, U.R. Acharya, F. Molinari, Cascaded LSTM recurrent neural network [50] K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition,
for automated sleep stage classification using single-channel EEG signals, Comput. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,
Biol. Med. 106 (2019) 71–81. 2016, pp. 770–778.
[37] P. Mamoshina, A. Vieira, E. Putin, A. Zhavoronkov, Applications of deep learning in [51] G. Huang, Z. Liu, L. Van Der Maaten, K.Q. Weinberger, Densely connected con-
biomedicine, Mol. Pharm. 13 (5) (2016) 1445–1454. volutional networks, Proceedings of the IEEE Conference on Computer Vision and
[38] H. Pratt, F. Coenen, D.M. Broadbent, S.P. Harding, Y. Zheng, Convolutional neural Pattern Recognition, 2017, pp. 4700–4708.
networks for diabetic retinopathy, Procedia comput. sci. 90 (2016) 200–205. [52] L.W. Jian, T.C. Lim, Automated detection of diabetes by means of higher order
[39] J. Pan, W.J. Tompkins, A real-time QRS detection algorithm, IEEE Trans. Biomed. spectral features obtained from heart rate signals, J. Med. Imag. Health Inf. 3 (3)
Eng. 32 (3) (1985) 230–236. (2013) 440–447.

10