Available online at www.sciencedirect.

com

ScienceDirect
ICT Express 4 (2018) 243–246
www.elsevier.com/locate/icte

Diabetes detection using deep learning algorithms

Swapna G. ∗, Vinayakumar R., Soman K.P.
Center for Computational Engineering and Networking (CEN), Amrita School of Engineering, Coimbatore, Amrita Vishwa Vidyapeetham, India

Received 4 September 2018; accepted 15 October 2018

Available online 8 November 2018

Abstract

Diabetes is a metabolic disease affecting a multitude of people worldwide. Its incidence rates are increasing alarmingly every year. If untreated,
diabetes-related complications in many vital organs of the body may turn fatal. Early detection of diabetes is very important for timely treatment
which can stop the disease progressing to such complications. RR-interval signals known as heart rate variability (HRV) signals (derived from
electrocardiogram (ECG) signals) can be effectively used for the non-invasive detection of diabetes. This research paper presents a methodology for
classification of diabetic and normal HRV signals using deep learning architectures. We employ long short-term memory (LSTM), convolutional
neural network (CNN) and its combinations for extracting complex temporal dynamic features of the input HRV data. These features are passed
into support vector machine (SVM) for classification. We have obtained the performance improvement of 0.03% and 0.06% in CNN and CNN-
LSTM architecture respectively compared to our earlier work without using SVM. The classification system proposed can help the clinicians to
diagnose diabetes using ECG signals with a very high accuracy of 95.7%.
⃝c 2018 The Korean Institute of Communications and Information Sciences (KICS). Publishing Services by Elsevier B.V. This is an open access
article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Keywords: Deep learning; Diabetes; Heart rate variability; ECG; CNN; LSTM

1. Introduction in heart rate. Thus, HRV is a marker to identify the presence of

Diabetes is a disease whereby blood sugar (glucose) is not neuropathy due to diabetes [1].
metabolized in the body. This increases the glucose in the blood Heart rate is the time interval between two consecutive
to alarmingly high levels. This is known by the name hyper- QRS complexes lying adjacent in ECG. The variation in RR
glycemia. In this condition, body is unable to produce sufficient interval is represented by HRV. The main attraction is that
insulin. The other possibility is that body cannot respond to the HRV measurement is non-invasive and reproducible [2]. A
produced insulin. Diabetes is incurable; it has to be controlled. variety of machine learning techniques has been proposed for
A diabetic person can develop severe complications like nerve the automated detection of diabetes in a non-invasive way.
damage, heart attack, kidney failure and stroke. According to Deep learning techniques, which can self-learn from data, have
statistics in 2017, an estimated 8.8% of global population has been increasingly employed for detecting diabetes now-a-days.
diabetes. This is likely to increase to 9.9% by year 2045. Conventional methodologies of feature selection and extraction
Hyperglycemia caused by diabetes, create abnormalities in are not required here.
the cardiovascular system independent of the possible presence In our present work, we analyse input HRV signals em-
of dyslipidemia, arterial hypertension etc. Diabetes causes
ploying deep learning architectures of CNN, LSTM and its
cardiovascular autonomic neuropathy (CAN) which completely
combinations. We achieve a high accuracy value of 95.7%
upsets the nervous system and results in diminished variability
employing CNN 5-LSTM architecture with SVM using 5-fold
∗ Corresponding author. cross-validation. This work is the sequel to our published earlier
E-mail address: swapna.goutham@gmail.com (Swapna G.).
Peer review under responsibility of The Korean Institute of Communica-
work making use of deep learning techniques in diabetes detec-
tions and Information Sciences (KICS). tion with HRV as input data achieving an accuracy of 95.1%.
https://doi.org/10.1016/j.icte.2018.10.005
2405-9595/⃝ c 2018 The Korean Institute of Communications and Information Sciences (KICS). Publishing Services by Elsevier B.V. This is an open access article
under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
244 Swapna G., Vinayakumar R., Soman K.P. / ICT Express 4 (2018) 243–246

Table 1 3.1. Recurrent neural network (RNN)

Summary of research works in diabetes detection with HRV data as input.
Authors Methods Accuracy Recurrent neural network (RNN) is capable of extracting
obtained (in %) dynamic temporal behaviour from an input time sequence.
Ref [4] Nonlinear 86.0 Basic RNNs are a network of nodes emulating neurons, each
Ref [5] Higher order spectrum 90.5 with a directed (oneway) connection to every other node. Each
Ref [6] Higher order spectrum 79.93
Ref [7] Nonlinear 90.0
node has a time varying real-valued activation. Each connection
Ref [8] Discrete wavelet transform 92.02 (synapse) has a real-valued weight which can be modified in
Ref [9] Empirical mode decomposition 95.63 every iteration. Nodes are either input nodes to receive data
Ref [3] Deep learning (CNN-LSTM) 95.1 from outside of the network or output nodes that yields results,
Proposed Deep learning (CNN-LSTM with SVM) 95.7 or hidden nodes that modify the data which passes through them
method via their route from input to output. The difference from the
traditional feedforward neural networks is that RNN is capable
of using its internal state, otherwise known as memory, to
The remaining part of the paper has the following orga- process sequences of inputs.
nization: Section 2 deals with previous important works for
automated non-invasive diabetes detection. Section 3 deals 3.2. Long short-term memory (LSTM)
with the crucial topic of deep learning and its architectural
variants. Dataset is described in Section 4. Section 5 displays Long short-term memory (LSTM) units are a special type
the proposed architecture. Section 6 gives information about of building units for RNN. It can analyse, classify and predict
experiments and results obtained. The paper concludes with temporal data sequences of time lags of any size. A typical
Section 6. LSTM network is made up of memory, input, output and
forget gates. The memory in LSTM can remember values over
2. Background topics and related works arbitrary time intervals. Each of the three gates is a form of
neuron (which computes an activation function of a weighted
A lot of research has happened on the non-invasive auto- sum). More than that, these gates control the passage of values
mated detection of diabetes using machine learning techniques. in LSTM layers; hence these special neurons are named as
Machine learning was employed based on steps of feature gates. By long short-term, the fact underlined is that LSTM’s
extraction, feature selection and classification. There were a memory can really last for large time duration. LSTM tackles
variety of works which differed in what type of features was the issue of exploding and vanishing gradient problem which is
extracted and what classifiers were tried upon. It was further an important issue while training traditional RNNs.
observed that the performance of traditional machine learning
algorithms is not up to the acceptable level in crucial arti- 3.3. Convolutional neural network (CNN)
ficial intelligence problems of speech recognition and object
recognition mainly because of the fact that the dimension Convolutional neural network (CNN) is an improvised vari-
of the data handled is high. The shortcomings of machine ant of multilayer perceptron. CNN is generally made up of an
input, an output layer and many hidden layers. The hidden
learning boosted the deep learning research. Deep learning also
layers of a CNN typically are made up of convolutional,
has its applications in healthcare. Lot of works has recently
pooling, and fully connected layers.
been published mainly in anomaly detection in the area of
healthcare. Related to diabetes detection, [3] used deep learning
3.4. Hybrid networks (CNN-LSTM)
techniques to detect diabetes from the input HRV data with
an accuracy value that closely matches with the maximum In hybrid networks, the initial part is CNN consisting of
accuracy achieved for automated diabetes detection till that convolution and maxpooling layers only. The maxpooling1D
date. In the proposed paper, we achieve the highest accuracy layer’s output is fed to the input layer of the next deep learning
value of 95.7% in diagnosing diabetes. Table 1 lists all the architecture like RNN or LSTM used.
important works on the automated non-invasive detection of
diabetes using HRV. 3.5. Support vector machine (SVM)

3. Deep learning In support vector machine (SVM), each data sample is

represented as a point in a space. It is ensured that a wide
Deep learning is a form of machine learning. Unlike in separation exists for samples of different categories. When a
machine learning, feature extraction and classification are not new data sample arrives, mapping to the space first happens.
explicitly done in deep learning networks. The hidden layers of The category of the new sample is decided depending on what
the deep learning network do all these implicitly within itself side of the dividing gap, the new data sample point lies. SVM’s
without involving the external researcher. A short description classification gap can be viewed as a hyperplane in case of
of deep learning networks is given below. binary classification. If more than two classes are present, then
 Swapna G., Vinayakumar R., Soman K.P. / ICT Express 4 (2018) 243–246 245

Fig. 1. (A) and (B) Sample heart rate signal for normal and diabetic subject.

the dividing gap can be viewed as a set of hyperplanes placed in

a high dimensional space. The optimal hyperplane is chosen in
such a manner that there is maximum possible distance from the
nearest sample on each side to the separating hyperplane. In our
case, classification is just to distinguish between the normal and
diabetic HRV, hence the basic binary SVM classifier is used.

4. Description of dataset
The Electrocardiograms (ECG) of 20 people each from
the diabetes and normal group were collected for 10 min
with people lying down in a relaxed supine position. The
heart rate time series data is derived from ECG signals using
Pan and Tompkins algorithm. This real-time algorithm can
effectively detect QRS complexes in an ECG signal based on
its morphological features like slope, amplitude and width. It
involves processes like digital bandpass filtering (to reduce
false detections due to noise) and thresholding operations (to
increase detection sensitivity). The ECG signal is sampled at
500 Hz. 71 datasets (same number for diabetic and normal
group) each were extracted from the recorded data. Each dataset
contains 1000 number of samples. The input data (Fig. 1) is Fig. 2. Overview of proposed architecture.
passed to deep learning algorithms without any further pre-
processing.
6. Experiments and results
5. Proposed architecture All experiments are run on GPU enabled TensorFlow [10]
An overview of proposed architecture is shown in Fig. 2. with Keras [11] framework. In this work, we use the same
This is composed of 3 main sections. In input layer, the heart configuration that we had used in our early paper [3]. In
rate variability of raw ECG signal are given as input to deep this work, we extract features in deep learning network, com-
learning architecture. This contains 5 CNN layers and each prised of CNN-LSTM architecture and pass into SVM for
layer follows maxpooling. First two CNN layers contain 64 and classification. LSTM has the capability to handle long-term
128 filters with filter length 3 and maxpooling with pooling dependencies in a data sequence. To decide the kernel function,
length 2. Next two CNN layers contain 256 and 512 filters we run two trail of experiment for SVM with linear and RBF
with filter length 3 and maxpooling with pooling length 4. A kernel. SVM with RBF kernel performed better. These SVM
model are implemented using Scikit-learn. The detailed 5-fold
last CNN layer contains 1024 filters with filter length 3 and
cross-validation accuracy is reported in Table 2. In almost
maxpooling with pooling length 6. This feature map is passed
all the network structures, SVM has performed better in 5-
into LSTM layer. LSTM contains 70 memory blocks which
fold cross-validation with accuracy which is comparable to the
learns the time domain features. This follows dropout 0.1.
fully connected linear with nonlinear activation function for
This randomly removes the neurons along with its connections. classification. Thus, we claim that the combination of SVM in
Finally, the features are passed into SVM for classification. The penultimate layer for classification with deep learning layers for
SVM used RBF kernel. This is defined as follows with samples feature extraction can achieve the best performance.
s and s1
( )
∥s − s1∥2 7. Conclusion and future work
K (s, s1) = exp (1)
2σ 2 Considerable part of human population is under the grip of
diabetes which is incurable. If not managed well, diabetes can
246 Swapna G., Vinayakumar R., Soman K.P. / ICT Express 4 (2018) 243–246

Table 2 Conflict of interest

Detailed results.
Architecture Accuracy obtained
The authors declare that there is no conflict of interest in this
paper.
CNN 1 with SVM 0.684
CNN 2 with SVM 0.755
CNN 3 with SVM 0.887 References
CNN 4 with SVM 0.913 [1] M.A. Pfeifer, D. Cook, J. Brodsky, D. Tice, A. Reenan, S. Swedine,
CNN 5 with SVM 0.939 J.B. Halter, D. Porte, Quantitative evaluation of cardiac parasympathetic
CNN 1-LSTM with SVM 0.743 activity in normal and diabetic man, Diabetes 31 (4) (1982) 339–345.
CNN 2-LSTM with SVM 0.764 [2] R.E. Kleiger, J.T. Bigger, M.S. Bosner, M.K. Chung, J.R. Cook, L.M.
CNN 3-LSTM with SVM 0.853 Rolnitzky, R. Steinman, J.L. Fleiss, Stability over time of variables
CNN 4- LSTM with SVM 0.937 measuring heart rate variability in normal subjects, Am. J. Cardiol. 68 (6)
CNN 5-LSTM with SVM 0.957 (1991) 626–630.
[3] G. Swapna, S. Kp, R. Vinayakumar, Automated detection of diabetes
using CNN and CNN-LSTM network and heart rate signals, Procedia
Comput. Sci. 132 (2018) 1253–1262.
lead to health hazards. Hence, early detection of diabetes is [4] U.R. Acharya, O. Faust, S.V. Sree, D.N. Ghista, S. Dua, P. Joseph, V.T.
extremely crucial. Nerve damages caused by diabetes, affect Ahamed, N. Janarthanan, T. Tamura, An integrated diabetic index using
the working of the heart. In the proposed work, HRV data is heart rate variability signal features for diagnosis of diabetes, Comput.
analysed to diagnose diabetes using deep learning techniques. Biomech. Biomed. Eng. 16 (2) (2013) 222–234.
[5] G. Swapna, U. Rajendra Acharya, S. VinithaSree, J.S. Suri, Automated
The maximum accuracy value of 95.7% was obtained for
detection of diabetes using higher order spectral features extracted from
CNN 5-LSTM with SVM network. This is the highest value heart rate signals, Intell. Data Anal. 17 (2) (2013) 309–326.
published for the automated diabetes detection with HRV as [6] L.W. Jian, T.-C. Lim, Automated detection of diabetes by means of higher
input data. Our non-invasive, flexible and reproducible system order spectral features obtained from heart rate signals, J. Med. Imaging
can serve as a reliable tool to clinicians to detect diabetes. Health Inf. 3 (3) (2013) 440–447.
[7] U.R. Acharya, O. Faust, N.A. Kadri, J.S. Suri, W. Yu, Automated
Further improvement in accuracy can be obtained using a very
identification of normal and diabetes heart rate signals using nonlinear
large sized input dataset. The potential of deep learning is measures, Comput. Biol. Med. 43 (10) (2013) 1523–1529.
so tremendous that it can take a big stride in future to the [8] U.R. Acharya, K.S. Vidya, D.N. Ghista, W.J.E. Lim, F. Molinari, M.
so far challengingly difficult area of anomaly prediction from Sankaranarayanan, Computer-aided diagnosis of diabetic subjects by
the anomaly detection if sufficiently large sized input data is heart rate variability signals using discrete wavelet transform method,
Knowl.-Based Syst. 81 (2015) 56–64.
available for research. The anomaly prediction can be tried
[9] R.B. Pachori, M. Kumar, P. Avinash, K. Shashank, U.R. Acharya, An
from the input data which may not have anomaly by extracting improved online paradigm for screening of diabetic patients using RR-
dynamic characteristics from the input data. The predicted interval signals, J. Mech. Med. Biol. 16 (01) (2016) 1640003.
information can serve as a warning signal for the patient as [10] M. Abadi, P. Barham, J. Chen, Z. Chen, A. Davis, J. Dean, M. Devin, S.
well as the doctor to take sufficient control and precautionary Ghemawat, G. Irving, M. Isard, et al., TensorFlow: A system for large-
scale machine learning, in: OSDI, vol. 16, 2016, pp. 265–283.
measures.
[11] F. Chollet, et al., Keras (2015).