Computer Methods and Programs in Biomedicine 154 (2018) 183–190

Contents lists available at ScienceDirect

Computer Methods and Programs in Biomedicine

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

Predictions of ocular changes caused by diabetes in glaucoma patients

Nicoleta Anton Apreutesei a,b, Filip Tircoveanu a,b, Alina Cantemir a,b, Camelia Bogdanici a,b,
Cătălin Lisa c,∗, Silvia Curteanu c,∗, Dorin Chiseliţă a,b
a
University of Medicine and Farmacy “Gr. T. Popa” Iasi, Surgery Department, Romania
b
Ophthalmology Clinic, University Street No 16, Iasi 700115, Romania
c
“Gheorghe Asachi” Technical University of Iasi, Faculty of Chemical Engineering and Environmental Protection “Cristofor Simionescu”, Department of
Chemical Engineering, Bd. Prof.dr.doc Dimitrie Mangeron No. 73, Iasi 700050, Romania

a r t i c l e i n f o a b s t r a c t

Article history: Background and objective: This paper builds different neural network models with simple topologies,
Received 18 April 2017 having one or two hidden layers which were subsequently employed in the prediction of ocular changes
Revised 1 November 2017
progression in patients with diabetes associated with primer open-angle glaucoma.
Accepted 14 November 2017
Material and Methods: For attempting to indicate whether there is a relationship between glaucoma and
diabetes, a simulation method, based on artificial neural networks (ANN), Jordan Elman networks (JEN)
Keywords: type, in particular, was applied in conjunction with clinical observation. The study was conducted on
Glaucoma a sample of 101 eyes with open angle glaucoma included and, in each case, the patients had associated
Diabetic retinopathy diabetes mellitus. A high degree of accuracy was exhibited by the models, demonstrating the potential ef-
Artificial neural networks
fectiveness of this artificial intelligence technique for predicting ocular changes associated with diabetes.
Direct and inverse modelling
The parameters considered in this study for modelling purpose were: glaucoma age, diabetes age, C/D
ratio (cup/disk size), glycated haemoglobin level (HbA1c), intraocular pressure (IOP), patient age, mean
deviation (MD) and LENS appearance.
Results: Relatively simple models, feed-forward neural networks with one or two intermediate layers,
provided clinically meaningful data in direct modelling, the probability of correct answers being of 95%.
Inverse modelling was also performed, in which MD depreciation was the output parameter. High accu-
racy was exhibited, in this case, with Jordan Elman networks, with the confidence interval of ±15%.
Conclusions: The neural models have demonstrated the possibility of their use in successfully predicting
the relationship between glaucoma and diabetes in a real clinical environment.
© 2017 Elsevier B.V. All rights reserved.

1. Introduction representative cases, with sufficient data from which it is able to

Diabetes and glaucoma are two chronic diseases that signifi-
cantly affect the population aged over 40; the opinion on the phys-
iopathology of these diseases when occurring in unison is not con-
sistent. Both diseases lead to degeneration of neural structures of
the eye. Many literature reports have found a correlation between
diabetes mellitus (DM) and primer open angle-glaucoma (POAG).
However, a nearly equal number of reports are unable to find evi-
dence of this correlation [1–4].
In order to establish whether there is a relationship between
glaucoma and diabetes, a simulation method based on artificial
neural networks (ANN) was used in conjunction with clinical ob-
servations. A neural network is capable of building its own algo-
rithm for solving a problem if it is trained by providing a set of
∗
Corresponding authors.
E-mail addresses: clisa@ch.tuiasi.ro (C. Lisa), scurtean@ch.tuiasi.ro (S. Curteanu).
extract the necessary information. The network is trained based on
collected data, i.e. process behaviour learning. In the work or refer-
ence phase, the network will use the acquired information to pro-
cess similar cases in the same way as those contained in the train-
ing set. Thus ANN can operate and draw conclusions from data not
available in the training process (“unseen data”) [5–7].
ANN have been widely used for data processing in medical di-
agnosis. Papik’s review shows that over 20 0 0 publications describ-
ing applications of neural networks in cardiology, oncology, pul-
monology, gastroenterology, radiology, ENT, gynaecology etc. were
published between 1990 and1997 [8]. Thanks to the aforemen-
tioned ability of ANN to predict outcomes based on limited train-
ing sets, they have found increasing utility as a diagnostic tool in
clinical practice. Besides their use in other specialties [9], ANNs
have been specifically used in ophthalmology, for measuring the
retinal structure and visual function. Various tools of artificial in-
telligence (neural networks and genetic algorithms) have been ap-

https://doi.org/10.1016/j.cmpb.2017.11.013
0169-2607/© 2017 Elsevier B.V. All rights reserved.
184 N. Anton Apreutesei et al. / Computer Methods and Programs in Biomedicine 154 (2018) 183–190
plied to determine visual field progression for patients with glau-
coma, to classify modifications in diabetic retinopathy (presence or
absence of retinopathy signs) and in the stage of glaucoma [10–17].
Antón et al. [18] used neural networks for interpretation of reti-
nal lesions in early perimetric glaucoma and concluded that both
neural networks and statistical analysis can accurately differenti-
ate early changes caused by glaucoma and other diseases, with an
accuracy of 97%. Other studies realized by Goldbaum show that
the networks differentiate better between normal and altered vi-
sual field than the overall visual field index. [19].
In recent studies, Rigla and others used artificial neural net-
works in the follow-up of patients with diabetes while promoting
the involvement of medical personnel in this area [20]. Further-
more, Singh et al. [21] used artificial intelligence tools in assess-
ing the risk of diabetic foot ulcers in patients with diabetes melli-
tus type by utilizing both multiple regressional analysis and neural
models. They concluded that the artificial networks have a better
predictive capacity than statistical models and could be used in as-
sessing the risk of foot ulcers in diabetic patients.
Inspired by previous excellent results in implementing ANN in a
diagnostic setting, ANN are used here for the first time to empha-
size the relationship between diabetes and glaucoma, as well as to
predict ocular modifications related to diabetes for patients with
glaucoma. By choosing diagnostic elements for the two diseases as
input parameters for building a training database (age glaucoma,
diabetes, report C/D, glycated haemoglobin, intraocular pressure,
LENS appearance, MD-deviation on visual field) and testing the re-
sulting ANN against a pool of patients not included in the training
set, a statistically significant correlation between the two disorders
was obtained.
2. Material and method
2.1. Study group
To predict the relationship between glaucoma and diabetes, a
database was constructed containing information on 52 patients
with glaucoma and diabetes, (101 eyes), examined in the Oph-
thalmology Clinic I of “ Sf. Spiridon” Hospital, in Iasi, between
2011 and 2013. The criteria taken into account in this study were:
normal or hypertensive primer open-angle glaucoma, ocular hy-
pertension and associated diabetes. Patients with other types of
glaucoma (pseudoexfoliation, pigment, cortisone, neovascular and
primer with angle closure) and those without associated diabetes
were excluded. Criteria for the diagnosis of normal and hyperten-
sive POAG were: age greater 35, IOP higher than 21 mm Hg without
treatment, camerular angle open to gonioscopy, optic nerve dam-
age typical in glaucoma (ratio C/D > 0.5), abnormal visual field (by
Humprey Field Analyzer perimetry) and retinal nerve fibre layer
with diffuse or localized defects (Optic Coherence Tomography -
Cirrus HD OCT Carl Zeiss Meditec) [22]. Assessment of diabetic
retinopathy modifications was made using Volk lens and by retinal
photographing on Fundus Camera Zeiss, on dilated pupil, following
their presence in all quadrants. Criteria for classification of changes
of DR (diabetic retinopathy) (ETDRS) were: no changes (absent
signs of DR), mild non-proliferative diabetic retinopathy (presence
of a single micro-aneurysm), mild form (micro-aneurysms, haem-
orrhages in 2–3 quadrants, venous dilatation and thornback rays
in one quadrant), severe (micro-aneurysms, haemorrhages in all
quadrants, venous dilatation in 2–3 quadrants) and proliferative di-
abetic retinopathy (retinal neovascularization on disk and retina in
different quadrants). Any new clinical element that appeared at
subsequent controls was considered as a progression of diabetic
retinopathy that led to its classification as severe nonproliferative
or proliferative diabetic retinopathy [23]. The information that led
Table 1
Characteristics of the patients included in the study.
Characteristics Average value
Ages 64.85 +/− 7.04
Sex 74.13% F, 25.87% M
Old glaucoma (years) 4.17 +/− 3.46
IOP 18.61 +/− 4.77
−4.63 db (medium)
Old diabetes (years) 8.05 +/− 6.08
Type diabetes
I 9.81%
II 58.92%
MIXT 31.37%
HgA1c (%) 6.98 +/− 1.24
Glycemic level(mg/dl) 137.42 +/− 23.17
DR changes Without DR 89.11%
Mild DR 5.94%
Moderate/severe DR 4.95%
to the database development included elements of diagnosis for
each disease (Table 1).
IOP (intraocular pressure), MD (mean deviation), type diabetes
(I - insulin, non insulin, mixt) DR (diabetic retinopathy), HgA1c
(glycated haemoglobin, marker of glycemic control, normal < 6,5%);
glycemic level (mg/dl) normal ( < 110 mg/dl).
2.2. Neural network design
Neural network modelling of the relation between glaucoma
and diabetes involves first the determination of the input param-
eters, as the main factors influencing this relation. Therefore, the
following measurements were selected as input parameters for
neuronal models: glaucoma age (A), diabetes age (B), C/D ratio
(C), glycated haemoglobin (D), intraocular pressure (E), patient age
(F), MD-mean deviation (H) and LENS appearance (G), which was
coded as 0 for non modifications and 1-for modifications. The con-
sidered output parameter was the presence or absence of diabetic
retinopathy modifications (DR). Regarding the output parameter
(O), “1” was used to mark the presence of modifications of DR and
“0” for no modifications.
Generally, in an inverse neural network modelling, the input
and output variables change their places. In order to reinforce the
idea of a correlation between diabetes and primer open- angle
glaucoma, in the last part of this study, reverse modelling was car-
ried out.
For inverse modelling, as output parameter was considered MD
(O”) and the following input variables: age glaucoma (A”), dia-
betes age (B”), C/D ratio (C”), glycated haemoglobin (D”), intraocu-
lar pressure (E”), the age of patients (F”), the presence or absence
of modifications of diabetic retinopathy - DR (G”) and the LENS ap-
pearance (H”).
Both types of neural modelling, direct and inverse, used com-
mon diagnosis elements for diabetes and glaucoma, trying to
demonstrate the existing relation between them. Mean deviation
is an important parameter for the two diseases, and was chosen as
the output in inverse modelling and as one of the inputs in direct
modelling.
2.2.1. Neural networks types
In order to find which topology is the best for neuronal models,
two types of neural networks, with different architectures, were
tested: feed forward neural network (MLP–multilayer perceptron)
and Jordan Elman Network (JEN) type.
Feed-forward neural networks are a class of networks com-
monly used in various fields, including medical diagnosis. They are
characterized by the presence of a layer of input neurons, a num-
ber of hidden layers and output layer neurons (Fig. 1). In Fig. 1,
 N. Anton Apreutesei et al. / Computer Methods and Programs in Biomedicine 154 (2018) 183–190
Fig. 1. Structure of a feed-forward neural network with two layers of hidden neu-
rons.
Fig. 2. Structure of Jordan Elman neural network.
several inputs are represented, I1 , I2 …In and a series of outputs,
O1 , O2 …Om ; intermediate layers has q and p neurons, respectively.
The benefit of this type of neural network consists, among other
things, in their ability to generalize, i.e. to operate well with data
sets which differ from those presented in the training phase.
The second type of network employed here is the Jordan Elman
Network, its representation being shown in Fig. 2. This type of net-
work extends the multilayer perceptron with context units which
are represented by processing elements which recall the previous
activity. Context units give to the network the ability to extract
temporal information from data, the Elman network copies the ac-
tivity from the first processing elements (first hidden layer) and
the Jordan network completes copying the output model. This type
of neural networks (recurrent neural networks) is also frequently
used in medical diagnosis [24–27]. Compared to other neural net-
works, JENs show high stability, outstanding performance in mak-
ing predictions and a higher similarity with human brain function-
ing because they benefit of “short term memory” through the pres-
ence of back loops.
2.2.2. Development of different types of neural networks
The data from the study group used to determine the neural
models was divided as follows: 79 were used in the training stage
of neural networks and 20 were kept for the testing stage. Two
of eyes from the original data (101) included in the study were
excluded due to the absence of some measuring parameters (i.e.
HgA1c value). So, 99 eyes remain available in the data set. The type
185
of algorithm which was chosen for training was a supervised train-
ing type - the back propagation algorithm. It is usually viewed as
the most important and widely used algorithm for training neural
networks. A specialized software NeuroSolutions was used to de-
velop and train neural networks. The number of training epochs
was 10 0,0 0 0 for all types of neural networks and the duration of
training ranged between 3 and 30 min, depending on the chosen
topology. This represents an acceptable time for obtaining a neural
model.
2.3. Statistical analysis
The resulted database was statistically analyzed using SPSS 18.0
and consisted in the enforcement of the following tests: ANOVA,
t-Student, F test, χ 2 test, ROC curve, Kaplan–Meier curve etc.
The statistical analysis of the data has the role of proving the
existence of a correlation between input parameters (age of glau-
coma, age of diabetes, ratio C/D, glycated haemoglobin, age of the
patients, mean deviation, LENS appearance) and output parame-
ters (DR/nonDR) that allows the construction of neural patterns
that correlate these parameters. Confidence intervals limits of 95%
were set and power (sensitivity) > 70% was considered for statisti-
cally relevant data using the established algorithm.
3. Results and discussions
3.1. Statistical analysis results
The subjects were divided into two groups depending on the
presence of diabetic retinopathy, as follows: DR Group-12 eyes
with diabetic retinopathy changes and non DR Group-89 eyes
without diabetic retinopathy changes.
The gender distribution highlights the increased frequency of
men with diabetic retinopathy. The age of the patients ranged
from 48 to 75 years and the average being 64.85 ± 7.04 years. In
eyes who presented diabetic retinopathy changes, the age of glau-
coma (Vgl) was significantly higher than the control group (5.67
vs 3.97 years; p = .049). The presence of glaucoma for more than
5 years induces an estimated risk of developing diabetic retinopa-
thy of about 1.5 times higher (RR = 1.48; IC95%: 1.02–1.75). All eyes
with diabetic retinopathy changes had diabetes for over 10 years:
66.7% men and 50% aged over 65 years. By marking the ROC curve
(Fig. 3), it is noted that the age of diabetes mellitus (Vdz) is a
good predictor of occurrence of diabetic retinopathy (AUC = 0.919;
IC95%: 0.854–0.985). Also, the age of glaucoma is correlated with
the manifestations of diabetic retinopathy with an influence of ap-
proximately 70% (AUC = 0.699; IC95%: 0.537–0.862), slightly lower
compared to the age of diabetes (Fig. 3).
With a variance of 20%, the values of HbA1c have varied
in the interval 5%–12.80%, the group average was 6.98 ± 1.38%,
without significant differences depending on the presence or ab-
sence of diabetic retinopathy changes (6.90% vs. 6.99%; p = .850).
Correlation between the age of diabetes and intra-ocular pres-
sure showed a strong indirect correlation in eyes with diabetic
retinopathy changes (r = −0.788; R2 = 0.6206; p = .002), approxi-
mately 79% of eyes with diabetic retinopathy changes associated
elevated intraocular pressure with reduced duration of diabetes.
Correlation between the age of diabetes and the mean devia-
tion of the visual field showed a strong indirect correlation in
eyes with diabetic retinopathy changes (r = −0.682; R2 = 0.4754;
p = .014), while for the nonDR group, the correlation was also indi-
rect, but of very low intensity (r = −0.011; R2 = 0.0 0 03; p = .918).
Correlation between the age of glaucoma and the mean devi-
ation of the visual field showed a very weak direct correla-
tion, both in eyes with diabetic retinopathy changes (r = +0.192;
R2 = 0.0341; p = .551) and in eyes without diabetic retinopathy
186 N. Anton Apreutesei et al. / Computer Methods and Programs in Biomedicine 154 (2018) 183–190
Fig. 3. The balance sensitivity/specificity of the age of diabetes and age of glaucoma in the determinism of diabetic retinopathy.
changes (r = +0.008; R2 = 0.0001; p = .941), insignificant statisti-
cally. In the cases studied, by marking the ROC curves, it is found
that the C/D (AUC = 0.700; IC95%: 0.565–0.835) and the mean de-
viation of the visual field (AUC = 0.770; IC95%: 0.657–0.863) enters
in the determinism of diabetic retinopathy with a weight of over
70%, while the level of haemoglobin (AUC = 0.420; IC95%: 0.245–
0.596) and intra-ocular pressure (AUC = 0.439; IC95%: 0.208–0.669)
were not good predictors of diabetic retinopathy changes (Fig. 4).
3.2. Neural network modelling results
Based on the statistical analysis of the database, the dataset
was found to be suitable for developing and training the neu-
ral models, correlating the output parameter - the presence or
absence of diabetic retinopathy changes - DR (O) with input
parameters–glaucoma age (A), diabetes age (B), C/D ratio (C), gly-
cated haemoglobin (D), intraocular pressure (E), patient age (F),
MD–mean deviation (H) and LENS appearance (G).
The next step in the neural network modelling was the creation
of data sets for training and validation: 79 eyes were used in the
training stage (meaning 80%) and 20 were retained for the testing
step (20%).
The number of hidden layers depends on the issue that will
be resolved. In general, a single hidden layer is sufficient to solve
most problems. However, for complex problems, up to three hid-
den layers may need to be used. Typically, the number of input and
output neurons corresponding to the number of input and output
variables, respectively, is determined by the nature of the applica-
tion. Neurons of hidden layers play an important role in detecting
regularities and rules contained in the training pattern. Often, the
way of determining the appropriate number of hidden neurons is
the trial and error method. Accordingly, several models were built
with eighth inputs, one or two layers of hidden neurons and one
output, in order to find the best topology for the neural model.
Table 2 shows some developed ANN topologies with their per-
formance, represented by the mean square error (MSE) (Eq. 1), the
correlation coefficient (r2 ) (Eq. 2) and percentage error (Ep ) (Eq. 3):
P N
j=1 i=1 (di j − yi j )2
MSE = (1)
N·P
where P is the number of the outputs (in this case, P = 1), N is
the amount of data, yij is the output value for i element with the
processing of elements j, and dij is the desired output for i with
 N. Anton Apreutesei et al. / Computer Methods and Programs in Biomedicine 154 (2018) 183–190 187

Fig. 4. Sensitivity/specificity balance of laboratory markers in the determinism of diabetic retinopathy.

Table 2 Analyzing the data presented in Table 2, it appears that excel-

Different ANN topologies.
lent results were obtained with all 10 neural models in the training
No. Network topology MSE r2 Ep .(%) Duration of training phase. With the ten candidate models in hand, the validation stage
stage (min) was carried out to check the response to the data that has not been
1. MLP(8:8:1) 0.0 0 0 0 09 0.999980 0.018 3.06 used in the training stage. Fig. 5 shows the results of the validation
2. MLP(8:16:1) 0.0 0 0 019 0.999964 0.015 3.50 step for some neural models from Table 2 where the experimen-
3. MLP(8:24:1) 0.0 0 0 0 06 0.999989 0.010 4.15 tal DR(O) was compared with the answer offered in the validation
4. MLP(8:32:1) 0.0 0 0 0 05 0.999990 0.0054 5.04
step by some neural models. For each neural model, standard de-
5. MLP(8:40:1) 0.0 0 0 0 04 0.999992 0.0106 5.38
6. MLP(8:16:8:1) 0.0 0 0 0 03 0.999995 0.0065 4.50 viation was calculated based on the following relationship (Eq. 4):
7. MLP(8:24:8:1) 0.0 0 0 0 02 0.999996 0.0083 6.29
8. MLP(8:24:16:1) 0.0 0 0 0 02 0.999996 0.0055 6.56

k 
 2
9. MLP(8:32:24:1) 0.0 0 0 0 01 0.999997 0.0029 9.20
10. MLP(8:40:32:1) 0.0 0 0 0 02 0.999996 0.0088 11.49 σ= (I )exp − (I )calc /(l − q ) (4)
i=1

where l is the amount of experimental data, and q is the number

processing of j. of parameters (in this case, q = 1). The results are given in Table 3.

 −
  −
 In Fig. 5, experimental values are compared with neural net-
Yexpi − Y exp · Yneti − Y net work results in the validation stage, which includes 20 instances
r2 =  (2)

 −
2  −
2 from a total of initial database composed of 99 data. Significant
Yexpi − Y exp · Yneti − Y net difference between experimental data and simulation results were
obtained for the patient with the code 19-TM. For the remain-
Yexp − Ynet ing patients, the output parameter O, calculated with the neu-
Ep = · 100 (3) ral model, is very close to the experimental values, which is also
Yexp
reflected by the values of the standard deviations (Table 3). The
where Y is the output data value and exp and net denote experi- best results and standard deviations were obtained with the neu-
mental values and those obtained from neural network models. ral model MLP(8:8:1). For all neural networks shown in Fig. 5, the
Network topology was encoded by ANN (m:n:r), particularly probability of a correct answer was 95% (19 correct answers from
MLP (m:n:r), where m is the number of neurons in the input layer, 20 possible).
n is the number of neurons in the hidden layer and r, the number The literature studies regarding the relationship between glau-
of neurons in the output layer. coma and diabetes, as well as the use of ANN in ophthalmol-
188 N. Anton Apreutesei et al. / Computer Methods and Programs in Biomedicine 154 (2018) 183–190

Fig. 5. Prediction of parameter O with different neural models in the validation stage. Patient eyes codes: 1- B. E.; 2-I. G.; 3- I. A.; 4-A. I.; 5-C. C.; 6-B. I.; 7-N. E.; 8-G. D.;
9-F. O.; 10-A. O.; 11- G. P.; 12- P. V; 13-C. A.; 14-I.G.; 15-C. F.; 16-C. C.; 17-C. E.; 18-C.G.; 19-T. M. and 20- C. F.

Table 3
Standard deviation values in the validation stage.

MLP (m:n:r) (8:8:1) (8:24:1) (8:32:1) (8:16:8:1) (8:24:8:1) (8:24:16:1) (8:32:24:1)

σ 0.024729 0.100825 0.089348 0.035991 0.050387 0.054259 0.123223

ogy, show that many of the authors are using neural networks Table 4
Different MLP and JEN topologies tested by inverse modelling.
built on the diagnostic glaucoma images or retinal diseases (visual
field, OCT). For the first time, in this manuscript, the data used No. Network topology MSE r2 Ep (%) Duration of the training
for the development of neural models are, in fact, the exact el- stage (min)
ements of diagnostic for both glaucoma and diabetic retinopathy. 1. MLP(8:8:1) 0.0 0 0466 0.997781 97.17 3.06
For direct comparison of our results with neural networks, multi- 2. MLP(8:16:1) 0.0 0 0245 0.998800 20.62 3.36
ple linear regression method (MLR) is also applied. This model was 3. MLP(8:24:1) 0.0 0 0217 0.998900 12.06 4.16
4. MLP(8:32:1) 0.0 0 0215 0.998900 9.33 3.07
obtained using the same database as for ANN model: 79 data were
5. MLP(8:40:1) 0.0 0 0214 0.998900 7.85 20.19
used to generate the model (training phase) and 20 dates were re- 6. MLP(8:48:1) 0.0 0 0215 0.998979 8.14 7.39
served for the validation stage. With MLR, the following relation 7. MLP(8:56:1) 0.0 0 0215 0.998974 11.03 7.35
resulted: 7. MLP(8:16:8:1) 0.0 0 0225 0.998920 13.00 23.49
8. MLP(8:24:8:1) 0.0 0 0220 0.998950 19.94 13.25
DR(O ) = − 0.069 + (0.00192 ∗ A ) + (0.0281 ∗ B ) − (0.188 ∗ C ) 9. MLP(8:24:16:1) 0.0 0 0214 0.998980 7.96 6.42
10. MLP(8:32:24:1) 0.0 0 0215 0.998978 9.50 17.43
− (0.0164 ∗ D ) + (0.00602∗E ) + (0.0 0 0525 ∗ F ) 11. MLP(8:40:32:1) 0.0 0 0216 0.998977 7.45 20.04
− (0.00545 ∗ G ) − (0.00491 ∗ H ) (5) 12. MLP(8:48:40:1) 0.0 0 0214 0.998980 7.32 16.03
13. MLP(8:56:48:1) 0.0 0 0214 0.998980 7.36 28.39
The relation (5) correlates the diabetic retinopathy changes - DR 14. JEN(8:16:1) 0.0 0 0 0 03 0.999998 5.08 5.22
15. JEN(8:24:1) 0.0 0 0 0 02 0.999998 1.37 6.31
(O) with parameters: glaucoma age (A), diabetes age (B), C/D ratio
16. JEN(8:32:1) 0.0 0 0 0 02 0.999999 1.15 7.21
(C), glycated haemoglobin (D), intraocular pressure (E), patient age 17. JEN(8:40:1) 0.0 0 0 0 01 0.999999 1.37 8.12
(F), MD - mean deviation (H) and LENS appearance (G). The model 18. JEN(8:48:1) 0.0 0 0 0 01 0.999997 2.89 8.56
was obtained using the same information as for the neural net- 19. JEN (8:16:8:1) 0.0 0 0 018 0.999912 35.66 6.33
work modelling: 79 eyes for the training phase. In the validation 20. JEN (8:24:16:1) 0.0 0 0 020 0.999951 10.18 8.28
21. JEN (8:32:24:1) 0.0 0 0 011 0.999949 13.16 9.40
stage, the probability of correct answers of the MLR model was
85% (17 correct answers from 20 possible), as in neural modelling
this probability was 95%. The standard deviation according to the
formula (4) for the DR prediction, in the validation stage for MLR, ing the output parameter with one of the input parameters. In in-
was calculated as 0.304826, a value much larger than the one ob- verse modelling, the output parameter was MD (O” ) and the fol-
tained from our neural models (Table 3). These results demonstrate lowing input variables were considered: glaucoma age (A”), dia-
that the neural model is more robust and accurate than other mod- betes age (B”), C/D ratio (C”), glycated haemoglobin (D”), intraocu-
els, particularly MLR. The same conclusion was obtained from the lar pressure (E”), patient age (F”), presence or absence of diabetic
studies of Kanhaiya and collaborators when they used artificial in- retinopathy changes - DR (G”) and LENS appearance (H”).
telligence tools to assess the risk of diabetic foot ulcers in patients The data which form the database was divided in training (79)
with diabetes mellitus type, using both multiple regressional anal- and validation (20) sets. Applying the trial and error method, sev-
ysis and artificial networks [21]. eral neural models with eight inputs were built, with one or two
To demonstrate the relationship between diabetes and primer intermediate layers of hidden neurons, and an output. Table 4
open-angle glaucoma, reverse modelling was carried out by replac- shows a series of neural networks with different topologies, of MLP
 N. Anton Apreutesei et al. / Computer Methods and Programs in Biomedicine 154 (2018) 183–190 189

4. Discussions

The artificial neural networks are an attempt to simulate the

structure and the functions of the brain that are specific to liv-
ing organisms. One of their most important characteristic is that a
model of the problem can be synthesized based on a great num-
ber of examples. With multiple applications in medicine, artificial
intelligence instruments found their use in ophthalmology in de-
tecting abnormalities in visual functions, thus playing a significant
role in diagnosing eye diseases.
Amato and his collaborators have recently published a literature
review in which they listed the main advantages of using various
artificial intelligence instruments in medical diagnosis, i.e. the pos-
sibility to process large data sets, the low probability of neglecting
relevant information, and reducing the time dedicated to establish-
ing diagnosis [27]. The database created in this study involves main
clinical elements for diagnosing glaucoma and diabetic retinopa-
thy. The statistical analysis results of the data obtained are simi-
Fig. 6. The results of the validation stage by inverse modelling. lar, or even better than those reported by other researchers in the
literature. After conducting the statistical analysis of the database,
Table 5
the creation of neural models was attempted. The most impor-
Standard deviation values in the validation stage for inverse modelling. tant stage in this process is the validation of networks, i.e. veri-
fying the responses they provide, compared to the data sets which
Model Type MLP (8:24:16:1) MLP (8:40:32:1) JEN (8:16:1) JEN (8:32:1)
have not been used in the training phase. In the direct and inverse
σ 3.171124 4.111157 5.198945 4.567930 modelling, performance of the neural model was rendered by the
mean square error (MSE) close to zero and the correlation coeffi-
cient (r²) close to 1. The probability for these neural networks to
Table 6 provide correct responses is of 95% (namely 19 out of 20 possi-
Predictions of the output parameter (MD) with the best neuronal model.
ble correct responses). In case of inverse modelling, MD deprecia-
Input data Output data MD (I’’) tion prediction respectively, if considering a ±15% confidence inter-
JEN(8:32:1) val, the JEN(8:32:1) model offers the best results, only two values
A’’ B’’ C’’ D’’ E’’ F’’ G’’ H’’
being located outside the confidence interval. The best model ob-
6 7 0.5 7 13 65 1 1 −13.3111
6 4 0.8 7.2 20 72 1 0 −1.3261
tained from validation in inverse modelling, namely JEN (8:32:1),
1 1 0.1 5 12 60 0 0 1.634219 was used in order to make predictions based on no experimen-
tal data. The output parameter offered by the neural model JEN
(8:32:1) was interpreted according to the input values. The output
values (large MD deviations) were positively correlated with the
and JEN types. Their performances, mean square error (MSE), cor- age of glaucoma and the degree of optic nerve damage (C/D ra-
relation coefficient (r2 ) and percentage error (Ep ) are presented in tio > 0.5) and the small MD values were positively correlated with
the same table. Better performance could be noticed in the training the incipient optic nerve changes or normal appearance. The ad-
stage for neural models of Jordan Elman (JEN) type. vantage of such neural networks is the possibility of offering very
Table S1 of the Supporting Information contains a comparison good precision in predicting glaucoma and retinal changes, using
between experimental data and the answers given by some of the as input the characteristic clinical features given by two conditions,
built neural models for the training phase. The neuronal responses in a relatively short time. The predictions made on the unfamil-
to the validation data are presented in Fig. 6, compared to the ex- iar data, which have shown good results in a relatively short time,
perimental values. For neural models whose results are shown in confirm the conclusion that different types of neural networks can
Fig. 6, standard deviation values were calculated with the relation effectively contribute to establishing a quick and accurate diagnosis
(4), in which, instead of O, O” was used (Table 5). In the valida- regarding to the changes, which occur in diabetic retinopathy. The
tion phase, the standard deviation of the two types of neural net- results prove that using artificial intelligence instruments can con-
works, MLP and JEN, were compared. If a ±15% confidence interval tribute to reducing the time for establishing the diagnosis, an issue
is considered, it was observed that the model JEN(8:32:1) provides highlighted by other researchers as well as in the review [27].
the best results with only two values being outside the confidence
interval. 4.1. Limits
The best validated neural model, JEN (8:32:1), was further used
to make predictions for the unfamiliar dataset. The input data is There are some limitations to this study-database size. A larger
presented in Table 6, which included clinical features of both dis- number of parameters related to patients with these conditions or
eases diabetes and glaucoma. It was found that the output values a higher database would provide more accurate estimates and ro-
(large deviation MD) were positively correlated with age glaucoma bust tests.
and optic nerve damage degree (C/D > 0.5) and lower MD at early
glaucoma and normal optic nerve. The advantage of using these 5. Conclusions
neural networks lies in the possibility of predicting with great ac-
curacy the changes in retinal and glaucoma, using as input given The neural models developed herein have demonstrated their
characteristic clinical features of the two diseases in a relatively usefulness in predicting the relationship between glaucoma and
short time. Therefore, the developed neural networks can effec- diabetes in a real clinical environment. The most accurate results,
tively contribute to establishing quickly and accurately diagnosis including low standard deviations, were those which predicted the
regarding changes of diabetic retinopathy. presence or absence of changes in diabetic retinopathy - DR, with
190 N. Anton Apreutesei et al. / Computer Methods and Programs in Biomedicine 154 (2018) 183–190
a feed forward neural model having one hidden layer and 8 neu-
rons, MLP(8:8:1). For this model, the probability to give correct an-
swers was 95% (i.e. 19 correct answers from 20 possible). For the
MLR model, the probability to give correct answers was 85% (17
positive answers from 20 possible). These results prove the neural
model is more robust and precise than model obtained with multi-
regression analysis.
In case of inverse neural network modelling, in order to prove
the relationship between glaucoma and diabetic retinopathy, with
a confidence interval of ±15%, it was found that the model
JEN(8:32:1) showed the best results, only two values being outside
the chosen confidence interval.
This study illustrates the fundamentals of building a neural
model suitable for predicting the rate and degree of modification
in optic change, when both glaucoma and diabetes are present.
Special emphasis is placed on the quality of models and their abil-
ity to reproduce, by simulation, new behaviour, which was un-
known in the training stage.
Conflict of interest
We wish to confirm that there are no known conflicts of inter-
est associated with this publication and there has been no signifi-
cant financial support for this work that could have influenced its
outcome.
Acknowledgment
This work was supported by the “Program 4, Fundamental and
Border Research, Exploratory Research Projects” financed by UEFIS-
CDI, project No. 51/2017.
Supplementary materials
Supplementary material associated with this article can be
found, in the online version, at doi:10.1016/j.cmpb.2017.11.013.
References
[1] M.O. Gordon, J.A. Beiser, J.D. Brandt, D.K. Heuer, E.J. Higginbotham, C.A. John-
son, J.L. Keltner, J.P. Miller, R.K. 2nd Parrish, M.R. Wilson, M.A. Kass, The ocular
hypertension treatment Study: baseline factors that predict the onset of pri-
mary open-angle glaucoma, Arch. Ophthalmol. 120 (2002) 714–720.
[2] R. Varma, M. Ying-Lai, B.A. Francis, B.B. Nguyen, J. Deneen, M.R. Wilson,
S.P. Azen, Los Angeles latino eye study group: prevalence of open-angle glau-
coma and ocular hypertension in latinos: the Los Angeles latino eye study,
Ophthalmology 111 (2004) 1439–1448.
[3] J.M. Tielsch, J. Katz, H.A. Quigley, J.C. Javitt, A. Sommer, Diabetes, intraocular
pressure, and primary open-angle glaucoma in the Baltimore eye survey, Oph-
thalmology 102 (1995) 48–53.
[4] A. Hennis, S.Y. Wu, B. Nemesure, M.C. Leske, Barbados eye studies group: hy-
pertension, diabetes, and longitudinal changes in intraocular pressure, Oph-
thalmology 110 (2003) 908–914.
[5] D. Dumitrescu, H. Costin, Teora Bucuresti (Eds.), Retele neuronale. Teorii si
Aplicatii, 1996.
[6] S. Curteanu, A. Dumitrescu, C. Mihailescu, B.C. Simionescu, The synthesis of
polyacrylamide-based multicompomnent hydrogels. A neural network mod-
elling, J. Macromol. Sci. Part A Pure Appl. Chem. A 46 (2008) 368–380.
[7] S. Curteanu, H. Cartwright, Neural networks applied in chemistry. I. Determi-
nation of the optimal topology of multilayer perceptron neural networks, J.
Chemometr. 25 (2011) 527–549.
[8] K. Papik, B. Molnar, R. Schaefer, Z. Dombovari, Z. Tulassay, J. Feher, Application
of neural networks in medicine-a review. diagnostic and medical technology,
Med. Sci. Monit. 4 (1998) 538–546.
[9] M. Huzum, D.V. Marţu, L. Rădulescu, S. Curteanu, Use of neural networks to
predict Eustachian tube permeability, J. Rom. Med. Dent. 14 (2010) 18–24.
[10] M.A. Simon, L. Alonso, A. Antón, A hybrid visual field classifier to support early
glaucoma diagnosis, Intel. Artif. Rev. Iberoam. Intel. Artif. 9 (2005) 9–17.
[11] G.G. Gardner, D. Keating, T.H. Williamson, A.T. Elliottet, Automatic detection on
diabetic retinopathy using an artificial neural network: a screening tool, Br. J.
Ophthalmol. 80 (1996) 940–944.
[12] C. Bowd, K. Chan, L.M. Zangwill, M.H. Goldbaum, T.W. Lee, T.J. Sejnowski,
R.N. Weinreb, Comparing neural networks and linear discriminant functions
for glaucoma detection using confocal scanning laser ophthalmoscopy of the
optic disc, Invest. Ophthalmol. Vis. Sci. 43 (2002) 3444–3454.
[13] H. Parsaei, M.H. Moradi, R. Parsaei, Development and verification of artificial
neural network classifiers for eye diseases diagnosis, 14th ICBME, 2008.
[14] K. Nouri- Mahdavi, D. Hoffman, D. Gaasterland, J. Caprioli, Prediction of vi-
sual field progression in glaucoma, Invest. Ophthalmol. Vis. Sci. 45 (2004)
4346–4351.
[15] M. García, C.I. Sánchez, M.I. López, D. Abásolo, R. Hornero, Neural network
based detection of hard exudates in retinal images, Comput. Methods Pro-
grams Biomed. 93 (2009) 9–19.
[16] A. Giangiacomo, D. Garway-Heath, J. Caprioli, Diagnostic glaucoma progres-
sion: current practice and promising technologies, Curr. Opin. Ophthalmol. 17
(2006) 153–162.
[17] G. Santos-García, E. Hernández Galilea, Using Artificial Neural Networks to
Identify Glaucoma Stages. Cap.18. The Mystery of Glaucoma. Edited by Dr.
Tomas Kubena, 2011.
[18] A. Antón, L. Zangwill, A. Emdadi, R.N. Weinreb, Nerve fiber layer measurements
with scanning laser polarimetry in ocular hypertension, Arch. Ophthalmol. 115
(1997) 331–334.
[19] M.H. Goldbaum, P.A. Sample, K. Chan, J. Williams, T.W. Lee, E. Blumenthal,
C.A. Girkin, L.M. Zangwill, C. Bowd, T. Sejnowski, R.N. Weinreb, Comparingma-
chine learning classifiers for diagnosing glaucoma from standard automated
perimetry, Invest. Ophthalmol. Vis. Sci. 43 (2002) 162–169.
[20] M. Rigla, G. García-Sáez, B. Pons, M.E. Hernando, Artificial intelligence method-
ologies and their application to diabetes, J. Diabetes Sci. Technol. (2017),
doi:10.1177/1932296817710475.
[21] K. Singh, V.K. Singh, N.K. Agrawal, S. Gupta, K. Singh, Association of Toll-like
receptor 4 polymorphisms with diabetic foot ulcers and application of artificial
neural network in DFU risk assessment in type 2 diabetes patients, Biomed.
Res. Int. (2013) 318686, 9 pages.
[22] European Glaucoma Society Terminology and Guidelines for Glaucoma IV-nd
edition 2014; Ch.2:2; pg. 79–89.
[23] International Council of Ophthalmology- ICO Guidelines for Diabetic Eye Care-
2014.
[24] A. Erfanian, B. Mahmoud, Real-time ocular artefact suppression using recur-
rent neural network for electro-encephalogram based brain–computer inter-
face, Med. Biol. Eng. Comput. 43 (2005) 296–305.
[25] X. Hu, H. Cammann, H.A. Meyer, K. Miller, K. Jung, C. Stephan, Artificial neu-
ral networks and prostate cancer—tools for diagnosis and management, Nature
Rev. Urol. 10 (2013) 174–182.
[26] A. Petrosian, D. Prokhorov, R. Homan, R. Dashei, D. Wunsch, Recurrent neural
network based prediction of epileptic seizures in intra- and extracranial EEG,
Neurocomputing 30 (20 0 0) 201–218.
[27] F. Amato, A. López, E.M. Peña-Méndez, P. Vaňhara, A. Hampl, J. Havel, Artificial
neural networks in medical diagnosis, J. Appl. Biomed. 11 (2013) 47–58.