P1: FLW/FGD P2: FLW

Journal of Medical Systems [joms] pp107-299073-5 March 6, 2001 18:43 Style file version Nov. 19th, 1999

Journal of Medical Systems, Vol. 25, No. 2, 2001

A Neural Network Approach in Diabetes Management

by Insulin Administration
G. Gogou,1 N. Maglaveras,1,4 B. V. Ambrosiadou,2 D. Goulis,3 and C. Pappas1

Diabetes management by insulin administration is based on medical experts’ experi-

ence, intuition, and expertise. As there is very little information in medical literature
concerning practical aspects of this issue, medical experts adopt their own rules for
insulin regimen specification and dose adjustment. This paper investigates the appli-
cation of a neural network approach for the development of a prototype system for
knowledge classification in this domain. The system will further facilitate decision
making for diabetic patient management by insulin administration. In particular, a
generating algorithm for learning arbitrary classification is employed. The factors par-
ticipating in the decision making were among others diabetes type, patient age, current
treatment, glucose profile, physical activity, food intake, and desirable blood glucose
control. The resulting system was trained with 100 cases and tested on 100 patient cases.
The system proved to be applicable to this particular problem, classifing correctly 92%
of the testing cases.
KEY WORDS: decision support; diabetes; insulin administration; neural networks.

INTRODUCTION

Insulin administration is particularly amenable to decision support system

development because it is highly based on a combination of “soft” knowledge ex-
pressed by the medical experts’ opinions, intuition and gut feel, as well as on “hard”
information such as laboratory results and other patient-related observations (symp-
toms and signs). The systems developed for diabetes management of type 1 and/or
2 diabetic patients by insulin administration can be classified into three general
categories: First, systems which model the biological functions of blood glucose/

1 Laboratory of Medical Informatics, Aristotle University of Thessaloniki, Greece.

2 Department of Computer Science, University of Hertfordshire, England.
3 Imperial College School of Medicine, St. Mary’s Hospital, London, UK.
4 To whom correspondence should be addressed at Laboratory of Medical Informatics, The Medical
School, Aristotle University of Thessaloniki, P.O. Box 323, GR-54006 Thessaloniki, Greece.

119
0148-5598/01/0400-0119$19.50/0 °
C 2001 Plenum Publishing Corporation
P1: FLW/FGD P2: FLW
Journal of Medical Systems [joms] pp107-299073-5 March 6, 2001 18:43 Style file version Nov. 19th, 1999

120 Gogou, Maglaveras, Ambrosiadou, Goulis, and Pappas

insulin relationship which are mathematical model based. These systems employ
such techniques as simulation, optimization, adaptive control and neural networks
to predict blood glucose response to insulin intake.(1,2) Second, systems that model
the decision-making process of experts involved in insulin administration. Such sys-
tems use symbolic manipulation such as expert system methodology(3,4) to prescribe
insulin at various dosages for “optimal” glucose control and they are called know-
ledge-based or expert systems. A third class of systems describes those that combine
mathematical and knowledge-based techniques and is called Intelligent Knowledge-
Based Systems. Almost all of the systems calculate insulin dose adjustment of one
insulin regimen or of a specific number of regimens.(2−7)
Neural networks have been used extensively in medical applications such as
image/signal processing,(8) as pattern and statistical classifiers(9) and for modeling
the dynamic nature of biological systems.(10)
Some of the systems employing neural networks are developed for decision
support purposes in diagnosis and patient management.(10,11) Such examples of neu-
ral network applications can be found in electrocardiology, in pattern recognition
either for QRS/PVC classification(8) or ischemia episode detection,(9) in diagnosis
and treatment of hypertension,(12) for knowledge classification and other.
The use of neural networks for diabetic patient management has also attracted
the interest of the medical informatics community because of their ability to model
the nonlinear dynamic nature of blood glucose control. Other neural network appli-
cations as applied in diabetes, for early diagnosis, are systems assisting knowledge
acquisition and expert system development combining knowledge-based and neural
network approaches.(13)
This paper presents the result of knowledge acquisition in the domain of insulin
regimen administration and describes a system developed for classifying the acquired
knowledge, based on a neural network (NN) approach. A generating algorithm is
employed for learning arbitrary classification.(14)

THE DOMAIN-INSULIN REGIME PRESCRIPTION

AND DOSE ADJUSTMENT

Insulin regimens and dose adjustment are prescribed by diabetologists depend-

ing on a number of factors such as diabetes type, patient age, activity during the day
and control targets. Medical experts depending on their own intuition, expertise and
experience on the subject of insulin administration consider these factors.
There are three insulin types depending on the beginning and duration of their
action. These are the fast, intermediate and long-acting insulin types. Insulin regimens
are insulin types in combinations administered in daily profiles. As insulin adminis-
tration proceeds food intake by approximately half an hour, it may take place before
breakfast, lunch, dinner, and bed.
The most widely used insulin regimens are:
1. short-acting insulin mixed with intermediate-acting insulin, given twice daily
before meals
P1: FLW/FGD P2: FLW
Journal of Medical Systems [joms] pp107-299073-5 March 6, 2001 18:43 Style file version Nov. 19th, 1999

Insulin Administration by Neural Network Decision Support 121

2. short-acting insulin mixed with intermediate-acting insulin, given before

breakfast AND short-acting insulin, given before evening meal AND inter-
mediate-acting insulin, given at bedtime
3. short-acting insulin, given three times daily AND intermediate-acting insulin,
given at bedtime
4. intermediate-acting insulin, given once daily

Generally, the regimens that contain long-acting insulin are not prescribed be-
cause they are considered outdated and they are used only out of necessity,(8−10)
especially in elderly patients. The intensive glucose control (regimen 3) gives the
best results in terms of blood glucose control but, on the other hand, results in
more hypoglycaemic episodes (low blood glucose) which is a, potentially, dangerous
situation. The result of intensive glucose control is the most frequent appearance of
hypoglycemia (low blood glucose) which is even more dangerous than hyperglycemia
(high blood glucose values).(15−17)
Depending on special occasions other regimens can also be prescribed. These
life circumstances include coexistence of another acute or chronic disease, short
life expectancy, honeymoon period, psychosomatic problems due to injections, the
environment and inability of the patient to understand the demands of a particular
regimen.(18)

THE GENERATING ALGORITHM FOR LEARNING

ARBITRARY CLASSIFICATION

The algorithm uses a three layer feed-forward neural network architecture. The
first, input, layer contains linear neurons, the second, hidden, layer contains sigma-pi
neurons, and the third, output, layer contains linear threshold neurons. The number
of neurons in the input and hidden layers equal the number of training patterns, and
the number of neurons in the output layer equals the number of classes (in our case
insulin regimens) that the neural network has to distinguish.(14)
The first layer has as input an n-dimensional pattern vector p ∈ Rn to be clas-
sified, along with a fixed reference number r ∈ R. Together p and r they form an
(n + 1)-dimensional vector ( p, r ) ∈ Rn+1 later mentioned as the extend input pat-
tern and denoted by p0 . The neurons’ outputs in the third layer are binary, so when
the output of the ith neuron is 1, the input pattern belongs to the ith class. In this
sense, the network output is defined to be the ordinal number of the neuron in the
output layer whose output is 1.
At the time of training, a set of training patterns { p1 , p2 , . . . , pnT } is given, where
nT is the number of training patterns. There is a set of classes {C1 , C2 , . . . , CnC },
where nC is the number of classes. Each training pattern belongs to one of the
classes, so we can define a mapping M such that for each training pattern p, M(p) is
the ordinal number of the class it belongs to, for example if pi belongs to C j then
M( pi ) = j.
P1: FLW/FGD P2: FLW
Journal of Medical Systems [joms] pp107-299073-5 March 6, 2001 18:43 Style file version Nov. 19th, 1999

122 Gogou, Maglaveras, Ambrosiadou, Goulis, and Pappas

THE PROTOTYPE SYSTEM FOR INSULIN REGIMEN

KNOWLEDGE CLASSIFICATION

The prototype system described in this paper classifies clinicians’ knowledge in

the domain of insulin administration. In this way, it may further support decision
making in this field. The present system has been designed so that the expert can
train it while it is in use. The whole study design is presented in Table I.

The System Knowledge

For the purposes of the development of the program for insulin regimen spec-
ification, a list of regimens was compiled by interviewing diabetes experts in the
United Kingdom and Greece (Step 1). Subsequently, a questionnaire was prepared
(Step 2) and was sent to three diabetic departments of UK and 14 diabetological
centers in Greece. The questionnaire was asking for the parameters that were nec-
essary in order to decide for each specific insulin regimen (Step 3), among the ones
that were proposed in Step 1. The outcome of the above process was a choice of
the main factors intuitively used by doctors for insulin regimen prescription. These
are:

• Diabetes type (unknown, type 1, type 2)

• Patient age
• What the patient is used to taking (unknown, tablets, insulin)
• Special condition (pregnancy, surgery, infection)
• Dawn phenomenon (yes, no)
• Unstable diabetes (yes, no)
• Glucose profile (morning, afternoon, evening, night/unknown, normal, hyper-
glycemia, hypoglycemia).
• Physical activity (morning-noon, afternoon-evening, night/none-unknown,
sedentary, light, heavy)
• Food intake (breakfast, lunch, tea, dinner)
• Desirable blood glucose control (fair, good, very good)

Table I. Study Design

Step No. Subject
1 Group of experts
in UK & Greece
2 Authors
3 Group of experts
in UK & Greece
4 Authors
5 Group of experts
in UK & Greece
6 Authors
Action
compilation of list: most widely-used
insulin regimens
compilation of questionnaire: decision-making
parameters for insulin regimen selection
completion of questionnaire: decision-making
parameters for insulin regimen selection
software development: Regimen Adjustment
completion of a table: 200 diabetic cases
complete with appropriate regimen
Neural Network training: training and testing
of Regimen Adjustment, using the 200 cases
P1: FLW/FGD P2: FLW
Journal of Medical Systems [joms] pp107-299073-5 March 6, 2001 18:43 Style file version Nov. 19th, 1999

Insulin Administration by Neural Network Decision Support 123

System Analysis
Knowledge-Based System Approach
Previous work carried out by the authors concetrated in the development of
a system that models the decision making of specialists in insulin administration
using knowledge based system techniques.(3,19,20) The system has been extensively
evaluated(20) and used for decision support in clinical practice.(19) Some of the as-
sumptions that simplified the prototype knowledge-based system development
proved to be significant limitations during the system’s evaluation. These were the
following:
1. The experts were considered to perform insulin administration following
a uniform decision-making model. This is not the case in reality. There is
a great deal of divergence among experts’ opinions on patient manage-
ment approaches, particularly in insulin regimen specification, depending on
many factors most important of which are their personal experience and the
patients’ lifestyles.
2. The system only considered that the expert had to choose from only a limited
number of options regarding insulin administration.
3. The system did not take into account important factors such as expected level
of blood glucose control, activity level, diet habits and other.
Limitations 2 and 3 above became evident during the clinical evaluation of the
system.(19) Limitations 1 and 3 were highlighted during the assessment phase with
respect to the level of agreement among a consensus specialists’ opinion and the
system.(20) What was required was a generic approach providing a versatile decision-
making model. This would consider a number of factors involved in patient manage-
ment for insulin regimen prescription. It would also be able to cope with any insulin
regimen that a diabetologist might wish to adopt.
The present work has identified these factors which total ten each one taking a
number of different values. An added difficulty was that the system would have to be
re-written every time new regimens were added since this would have involved new
inferences among insulin types administered in combinations, blood glucose/insulin
values together with the increasing complexity of the additional factors participating
in the decision making process as discussed in section 4. The System Knowledge
earlier.
The Neural Network Approach
The neural network approach employed in this paper overcomes the above
mentioned difficulties. It further provides a generic tool for classifying existing as
well as new knowledge in insulin regimen administration. Previous work in applying
a neural network approach was based on a backpropagation algorithm and a limited
number of regimens.(11) The convergence to a local optimum depended on the initial
condition chosen and there was no guarantee whether the system would arrive at
an optimum altogether. We can not carry out a direct comparison of the algorithmic
techniques employed, because the same number of factors, regimens and patient
cases have not been considered in the two neural network approaches.
P1: FLW/FGD P2: FLW
Journal of Medical Systems [joms] pp107-299073-5 March 6, 2001 18:43 Style file version Nov. 19th, 1999

124 Gogou, Maglaveras, Ambrosiadou, Goulis, and Pappas

Fig. 1. Regimen adjustment.

The System Interface

The main objectives of the user-interface design are the increase of the speed
of learning and speed of use, the encouragement of rapid recall of how to use the
interface and the increase of the attractiveness to the potential users. Every data
subject was given a separate window form and several fields were grouped through
menu driven input. Icons and mouse input were also used in order to minimise
complicated keyboard requirements.
In order to suggest the appropriate insulin regimen the user has to enter the com-
bination of parameters for the patient (Fig. 1), given under The System Knowledge,
and asks the program for its “Advise.” The system uses these parameters as input
to the neural network, as described under The Generating Algorithm for Learning
Arbitrary Classification, which outputs a suggested regimen (Fig. 2).

Fig. 2. Suggested regimen.

P1: FLW/FGD P2: FLW
Journal of Medical Systems [joms] pp107-299073-5 March 6, 2001 18:43 Style file version Nov. 19th, 1999

Insulin Administration by Neural Network Decision Support 125

Fig. 3. Addition of new pattern.

If the regimen suggested by the system is not acceptable to the expert, he can
add the parameters entered along with his regimen suggestion, as a new training
pattern (Fig. 3). The main advantage of this system is that every expert can enter his
own regimen suggestions so that the neural network can be trained according to his
point of view or experience.
This is the reason why this particular neural network algorithm was chosen
instead of another. The neural network can be trained initially by the expert and
continue training while it is in use. In other algorithms (back-propagation, for exam-
ple) the network should already be trained and any objection to its decisions means
that it should be trained all over again.

RESULTS

For the purposes of system training in Insulin Adjustment, a number of diabetol-

ogists came up with the most common regimens used (Step 1) (Table II) as well as 200
cases of diabetic patients, complete with the selection of the appropriate regimen,
according to their opinion. (Step 5) (Table III). The 200 cases were seperated into
two set, 100 cases were used for the training of the neural network and the next 100
for testing it. The explanation of the parameters in Table III can be found in Table IV.
For example:
Case No. 1 shows a typical case of IDDM: child with moderate activity. Our
target (optimal blood glucose control in order to avoid late complications) demands
the use of multiple-dose regimes (3–4 daily). The regime suggested consists of 4 shots
and provides the best solution for blood glucose optimization.
Case No. 8 is a case of NIDDM: elderly with a medical condition that suggests
the use of multiple-dose regimes (4–6 daily) of short acting insulin. Our target is
the optimal blood glucose control in order to avoid acute complications. The regime

Table II. Most Common Regimens

Regimen No. Pre-Breakfast Pre-Lunch Pre-Tea Pre-Bed
1 Intermediate acting — — —
2 Short + intermediate — Short + intermediate —
acting — acting —
3 Short acting Short acting Short acting Intermediate acting
4 Short acting Short acting Short acting Short acting
 P1: FLW/FGD
P2: FLW
Journal of Medical Systems [joms]
pp107-299073-5

Table III. Sample of Training Patterns

Case DM Previous Regimen
No. type Age Special Rx Target Dawn Unstable BG-bre BG-lun BG-din BG-bed PA-mor PA-aft PA-nig FI-bre FI-lun FI-din FI-bed No.
1 1 20 n i vg n n n hi hi hi hi l — y y y y 4
March 6, 2001

2 1 20 n i vg n n nl nl nl nl h — — y y y y 3
3 1 35 n i vg y n lo nl nl nl s — — y n y y 2

126
4 1 25 n i vg y n hi lo nl hi — — h y n n y 2
5 1 45 n i g y y hi hi hi hi s — — y y y y 3
18:43

6 1 50 n i g n n nl nl nl nl h h — y n y n 2
7 2 40 y i vg n y hi hi hi hi h — — y y y y 4
8 2 80 y i vg n n hi hi nl nl s — — y y y y 4
9 2 45 n t vg n n nl nl nl nl s s — y y y y 3
10 2 50 n t vg y n hi lo nl lo s — — y n y n 2
Style file version Nov. 19th, 1999
P1: FLW/FGD P2: FLW
Journal of Medical Systems [joms] pp107-299073-5 March 6, 2001 18:43 Style file version Nov. 19th, 1999

Insulin Administration by Neural Network Decision Support 127

Table IV. Explanation and Value Ranges of the Categories of Table III
Parameter Values Explanation
DM type Type 1 (1), type 2 (2) Diabetes mellitus type
Age Number Age in years
Special case Yes (y), no (n) Special condition: pregnancy,
surgery, infection
Prev. Rx Insulin (i), tablets (t) Patient’s previous regime
Target Fair (f), good (g), very good (vg) Desirable diabetes control
Dawn Yes (y), no (n) Dawn phenomenon
Unstable Yes (y), no (n) Unstable diabetes
BG-bre Normal (nl), hyperglycemia (hi), hypoglycemia (lo) Blood glucose - breakfast
BG-lun Normal (nl), hyperglycemia (hi), hypoglycemia (lo) Blood glucose - lunch
BG-din Normal (nl), hyperglycemia (hi), hypoglycemia (lo) Blood glucose - dinner
BG-bed Normal (nl), hyperglycemia (hi), hypoglycemia (lo) Blood glucose - bed
PA-mor Sedentary (s), light (l), heavy (h) Physical activity - morning
PA-aft Sedentary (s), light (l), heavy (h) Physical activity - afternoon
PA-nig Sedentary (s), light (l), heavy (h) Physical activity - night
FI-bre Yes (y), no (n) Food intake - breakfast
FI-lun Yes (y), no (n) Food intake - lunch
FI-din Yes (y), no (n) Food intake - dinner
FI-bed Yes (y), no (n) Food intake - bed

suggested consists of 4 shots and provides the best solution for blood glucose opti-
misation under these conditions.
Case No. 10 is a typical case of NIDDM: adult with low physical activity. Our
target (optimal blood glucose control in order to avoid acute and late complications)
demands the use of multiple-dose regimes (2 shots daily). The regime suggested
consists of 2 shots of intermediate acting insulin and provides a good solution for
blood glucose control.
The 100 cases were entered as training patterns to the NN. The next 100 cases
were used for testing the system. The overall performance of the NN can be seen in
Tables V to IX.

DISCUSSION

Differences Between Systems Utilizing the Generating and

Back-Propagation Algorithms
There can not be a direct comparison of results in terms of insulin administra-
tion, between the work the neural network algorithm described in this paper and

Table V. Testing Results

NN regimen
1 2 3 4 %
Correct 1 12 0 0 0 100
regimen 2 4 28 4 0 78
3 0 0 38 0 100
4 0 0 0 14 100
Total 92
P1: FLW/FGD P2: FLW
Journal of Medical Systems [joms] pp107-299073-5 March 6, 2001 18:43 Style file version Nov. 19th, 1999

128 Gogou, Maglaveras, Ambrosiadou, Goulis, and Pappas

Table VI. Testing Results for Regimen No. 1

Regimen 1 Decision Model
+ − Sensitivity
Truth + 12 0 12 100%
− 4 84 88 Specificity
16 84 100 96%

Table VII. Testing Results for Regimen No. 2

Regimen 2 Decision Model
+ − Sensitivity
Truth + 28 8 36 78%
− 0 64 64 Specificity
28 72 100 100%

Table VIII. Testing Results for Regimen No. 3

Regimen 3 Decision Model
+ − Sensitivity
Truth + 38 0 38 100%
− 4 58 62 Specificity
42 58 100 94%

Table IX. Testing Results for Regimen No. 4

Regimen 4 Decision Model
+ − Sensitivity
Truth + 14 0 14 100%
− 0 86 86 Specificity
14 86 100 100%

previous work employing the back-propagation algorithm.(11) This is due to the dif-
ferent parameters involved in the decision-making process.
Back-propagation is essentially a gradient descent search in the weight space
to minimize an error function. As with other algorithms employing the gradient
descent search strategy, there exists the possibility that the algorithm gets stuck in a
local minimum. In other words, it is not guaranteed to converge to the global solution
although it reportedly converges in most cases. The algorithm is slow, it can take hours
for a moderate-sized network to converge on a fast PC. Back-propagation is sensitive
to its initial condition,(21−23) generalization of the trained network depends on the
choice of a set of initial conditions such as the network size and initial random weights.
Overgeneralization is also reported.(24) It is generally difficult to stop training at a
point where generalization is good.
Generating algorithm, on the other hand, achieves 100% correct classification
rate on training patterns. The algorithm is very fast, only 50 msec were required to
train (build) the netwrok. This algorithm, however, is applicable in cases where the
number of classes is limited and manageable by the trainer.
P1: FLW/FGD P2: FLW
Journal of Medical Systems [joms] pp107-299073-5 March 6, 2001 18:43 Style file version Nov. 19th, 1999

Insulin Administration by Neural Network Decision Support 129

The Neural Network Approach in Insulin Regime Selection

The Neural Network approach, although a black box technique, appears to
be more suitable than the knowledge-based one, in that it may encapsulate the
ever expanding knowledge of experts in this domain. There is only a handful of
rules that have almost universal application, such as “If a special condition, like
surgery or severe acute illness, is present, then use a regimen that contains short-
acting insulin only.” In the vast majority of the cases, the decision concerning in-
sulin regimen selection comes after the simultaneous weighting of the parame-
ters that have been identified by our study and are presented under The System
Knowledge.

Methodological Issues
Throughout the study, we used groups of experts from Greece and UK (Steps 1,
3, and 5). This involvement was much more preferable than knowledge acquisition
through simple literature review, as it is well recognized that insulin regimen selection
is a problem that has more than one scientifically acceptable solutions. In Step 1,
diabetologists prepared lists of personally preferred insulin regimens and the most
widely used of them compiled a master list. In the same way, in Step 3, the experts
were asked for the necessary input in order to decide, in general, that is not for a
specific regimen, about insulin scheme selection. Finally, in Step 5, we compiled a
table that included some possible combinations of the selected parameters in Step 3.
We did not aim specifically for “grey” diabetic cases, not least because we were not
sure which the “grey” cases were. We thought that if we provide the experts with a
large enough number of cases, all trends would be adequately represented. For this
study, we just used the more widely selected regimen among the experts in each case,
as input for the NN. For a more detailed convergence in expert opinions a Delphi
approach is required.(20)

CONCLUSIONS

The Insulin Dependent Diabetes Mellitus Advisor developed was an effort to

systematize the problem of Insulin Regimen Prescription. Ten factors were consid-
ered assigning to each one of them different values.
The neural network can be trained initially by its user (diabetologist) and
continue training while it is in use with little effort and without significant loss of
time. In this way, the system overcomes any variation arising due to the differences
in patient life style and diet habits as well as varying expert opinions in this field. The
rather small number of classes that we have makes the continuous training process
a realistic possibility.
It must be remembered that insulin prescription is a problem that has more than
one solutions. Various medical experts may suggest a variety of scientifically correct
P1: FLW/FGD P2: FLW
Journal of Medical Systems [joms] pp107-299073-5 March 6, 2001 18:43 Style file version Nov. 19th, 1999

130 Gogou, Maglaveras, Ambrosiadou, Goulis, and Pappas

solutions. Therefore, the contribution of the generating algorithm lies in the fact that
it may take these differences of opinions into account.

ACKNOWLEDGMENT

This work was supported in part by the HC-1010 DIABCARD-3 project from
the CEC.

REFERENCES

1. Lehmann, E. D., and Deutsch T., AIDA: A MK II Automated Insulin Dosage Advisor. J. Biomed.
Eng. 15:201–242, 1993.
2. Sano, A., Adaptive and optimal schemes for control of blood glucose levels. Biomed. Meas. Inf. Contr.
1:16–22, 1986.
3. Ambrosiadou, V., Alevizos, M., and Ziakas, G., Decision support in diabetes management for optimal
glycaemic control. IEEE Proc Systems Man and Cybernetics, Systems Engineering in the Service of
Humans. Decision Making 5:391–397, 1993.
4. Schneider, J., Piwernetz, K., Engelbrecht, R., and Renner, R., DIACONS-A consultation system to
assist in the management of diabetes; DIAMON- An expert system to assist in the therapy of diabetes.
Expert Systems and Decision Support in Medicine, (Springer, Heidelberg), pp. 44–49, 575, 1988.
5. Chanoch, L. H., Jovanovic, L., and Peterson C. M., The evaluation of a pocket computer as an aid to
insulin dose determination by patients. Diabetes Care 8:172–176, 1985.
6. Permick, N., and Rodburd D., Personal computer programs to assist with home monitoring of blood
glucose and self-adjustment of insulin dosage. Diabetes Care 9:61–69, 1986.
7. Dimitrov, A., Nestorov, I., and Christov, V., A two-stage adaptation scheme for computer-aided
adjustment of intensified insulin dosage regimens. World Congress of Medical Physics and Biomedical
Engineering Proceedings, PS17-1.10, 1994.
8. Miller, A. S., Blott, B. H., and Hames, T. K., Review of neural network applications in medical imaging
and signal processing. Med. Biol. Eng. Comput. 30:449–464, 1992.
9. Maglaveras, N., Stamkopoulos, T., Pappas, C., and Strintzis M., An adaptive back-propagation neural
network for real-time ischemia episodes detection. Development and performance analysis using the
European ST-T database. IEEE Trans. Biomed. Engng. 45(7):805–813, 1998.
10. Lakatos, G., Carson, E. R., and Benyo, Z., Artificial neural network approach to diabetic management.
Proceedings of the Annual International Conference of the IEEE, EMBS, 1010–1011, 1992.
11. Ambrosiadou, V., Gogou, G., Pappas, C., and Maglaveras, N., Decision support for insulin regime
prescription based on a neural network approach. Medical Informatics 21(1):23–34, 1996.
12. Poli, R., Gagnoti, S., Livi, R., Goppini, G., and Vali, G., A neural network expert system for dia-
gnosing and treating hypertension. IEEE Computer 24(3):64–71, 1991.
13. DeClaris, N., and Su, M. C., A neural network based approach to knowledge acquisition and expert
systems. IEEE Systems Man and Cybernetics Proc. 2:645–650, 1991.
14. Chen, Y., Thomas, D., and Nixon, M., Generating-shrinking algorithm for learning arbitrary classifi-
cation. Neural Network 7(9):1477–1489, 1994.
15. Foster, D. W., Diabetes mellitus. In Harrison’s Principles of Internal Medicine, (11th edition),
McGraw-Hill, pp. 1778–1797, 1987.
16. Garber, A. J., Diabetes mellitus. In (J. H. Stein, ed.), Internal Medicine (4th edition), Mosby, pp. 1391–
1430, 1994.
17. World Health Organization (Europe) and International Diabetes Federation (Europe), Diabetes
Care and Research in Europe, The St Vincent Declaration, Diabetic Med, 1990.
18. Group of European Politics IDDM, Instructions for the Treatment of Insulin-Dependent (type I)
Diabetes-Consensus Guidelines, 1993.
19. Ambrosiadou, V., Goulis, D. G., and Pappas, C., Clinical Evaluation of the DIABETES Expert
System for Decision Support by Multiple Regimen Insulin Dose Adjustment. Computer Methods
and Programs in Biomedicine, 49:105–115, 1996.
P1: FLW/FGD P2: FLW
Journal of Medical Systems [joms] pp107-299073-5 March 6, 2001 18:43 Style file version Nov. 19th, 1999

Insulin Administration by Neural Network Decision Support 131

20. Ambrosiadou, V., The DELPHI Method as a Consensus and Knowledge Acquisition Tool for the
Evaluation of the DIABETES System for Insulin Administration, Medical Informatics and the In-
ternet in Medicine, 24(4):257–268, 1999.
21. Masters, T., Practical Neural Network Recipes in C++, Academic Press, London, 1993.
22. Hart, A., and Wyatt, J., Evaluating black-boxes as medical aids: Issues arising from a study of neural
networks. Med. Inf. 15(3):229–236, 1990.
23. Kolen, J. F., and Pollack, J. B., Back-propagation is sensitive to initial conditions. Complex Systems
4:269–280, 1990.
24. Hall, C., and Smith, R., Pitfalls in the application of neural networks for process control. IEEE
Control Engineering Series 46: Neural Networks for Control and Systems 243–256, 1992.