AIDA Research Use
(Re)-Load AIDA Frames / Menus
In addition to its possible educational / teaching / self-learning / demonstration uses AIDA may also be able to support research work into the application of computers in diabetes care. Please use the menu below to review some research projects that have involved AIDA. (This section of the Website highlights areas particularly where other research groups have wanted to make use of simulated blood glucose data in their own research work).
- Testing decision support prototypes
- Qualitative models and fuzzy systems: an integrated approach for learning from data
- Learning from data through the integration of qualitative models and fuzzy systems
- How to improve fuzzy-neural system modeling by means of qualitative simulation
- Adaptive controllers for intelligent monitoring
- A distributed system for diabetic patient management
- Protocol-based reasoning in diabetic patient management
- Testing decision support prototypes #2
- Testing decision support prototypes #3
- Artificial Neural Networks
- Artificial Neural Networks - University of Strathclyde Prototypes
- Artificial Neural Networks - University of Strathclyde - Prototype #2 - Study #1
- Artificial Neural Networks - University of Strathclyde - Prototype #2 - Study #2
- Artificial Neural Networks #2
- Artificial Neural Networks #3
- Dietary assessment
- Diabetes model - further developments
- Web based diabetes simulator: AIDA on-line
- On-going research using AIDA
- Further ideas - generating simulated blood glucose data - for training / validation
- Learned space of parameter interactions using simulated blood glucose data
- Performance monitoring of closed-loop insulin delivery devices
- A business plan utilising telehealth-care technology within the home
- Evaluation / validation research usage of AIDA
- Other research uses of AIDA / Placing technical information on the Web
- Have a research project of your own?
- Research publications on-line
Testing decision support prototypes #1:Bellazzi and colleagues from the Dipiartimento di Informatica e Sistemistica, University of Pavia in Italy have described using simulated blood glucose data from AIDA to test out a number of computer decision-support prototypes under development in their laboratory. In Dr. Bellazzi's own words:
"I started working on the use of information technology tools in diabetesmellitus in 1994, and, due to my background in engineering, one of my firststeps was to understand the medical problem and its patho-physiology. At thesame time I also studied papers that could add a mathematical basis to themedical concepts I was learning. During these activities I read papers aboutthe AIDA model and downloaded the AIDA program that, at that time, was runningunder the DOS operating system.
Since that period I have exploited the diabetes simulator implemented inAIDA in a large number of research projects, and the model was used as thebasis for several Masters theses [1-3] at the School of Engineering, where I am teaching. Moreover, AIDA was used by one of my co-workers at the Departmentof Pediatrics of the Policlinico S. Matteo Hospital in Pavia.
To mention some of the more interesting activities we did, we exploited theAIDA model for:
- Testing of different strategies (algorithms) for insulin optimization inroutine care. In this area we compared different approaches, fromrule-based systems, to Fuzzy-rule based systems and finally to model-basedstrategies.
- Testing of a new model, based on Fuzzy systems and qualitative modeling,for blood glucose forecasting in patients with type I diabetes mellitus. Inthis area we used the patient simulator to obtain data and test thepredictions in different situations and insulin protocols.
- Education of Masters students of the school of engineering about diabetes.To this end we used AIDA to quickly introduce diabetes to people that wantedto Master in diabetes modeling.
Moreover, together with the University of Padova, it was decided to use AIDAfor comparison in the development of a new simulator of type I diabetesmellitus within the European Union IV Framework funded project, calledT-IDDM (Telematic Management of Diabetes Mellitus patients).
As a researcher, I am therefore in debt to the AIDA model for its help in myresearch activities, that fostered new approaches and ideas. As a teacher, Ibelieve that it is a valuable instrument for teaching students that do notcome from the school of medicine and who quickly need to learn aboutdiabetes regulation mechanisms."
Return to the Menu
Qualitative models and fuzzy systems: an integrated approach for learning from dataIn their report entitled "Qualitative models and fuzzy systems: an integrated approach for learning from data" Bellazzi and colleagues  have described a prototype method for the identification of the dynamics of non-linear systems in diabetes care by trying to learn from data. The key idea which underlies their approach consists of the integration of qualitative modelling techniques with fuzzy logic systems. The resulting hybrid method exploits the a priori structural knowledge of the system to initialise a fuzzy inference procedure which determines, from the available experimental data, a functional approximation of the system dynamics that can be used as a reasonable predictor of the patient's future state. The major advantage that is believed to result from such an integrated framework lies in a significant improvement in both the efficiency and robustness of identification methods based on fuzzy models which learn an input-output relation from the data provided. As a benchmark for the methodology, the authors have considered the problems of identifying the response to insulin therapy for insulin dependent (type 1) diabetic patients.
For this work, Bellazzi et al , used simulated data from AIDA, in which the noise level could be suitably manipulated for test purposes, as well as the number of missing data varied / increased. The overall procedure occurred in two steps:
(1) Training phase: Over a simulated period of 24 hours, with different sampling times ranging from 15 minutes to 8 hours, the authors have simulated patient responses to an injection of regular insulin followed by a meal. The performance of a function approximator, y(x), obtained using the prototype has been compared, through root mean square error (RMSE) calculations, with that of a separate function approximator identified just from the data.
(2) Validation phase: Evaluation of the predictive accuracy of the prototype's function approximator [y(x)] has taken place when dealing with a new dataset. In particular the new data was obtained by simulating the patient response to a typical daily insulin protocol, composed of two injections of NPH insulin and two injections of regular insulin, followed by a meal.
Bellazzi and colleagues have tested both phases in different experimental settings . (i) Training and validation have taken place with noise-free data and a minimum sampling time (of 15 mins); (ii) Training and validation with noisy data and a minimum sampling time; and (iii) Training and validation with a maximum sampling time.
The preliminary results obtained using AIDA simulation data have motivated the authors to do further work with their approach, moving towards an evaluation of the method with real patient data.
Return to the Menu
Learning from data through the integration of qualitative models and fuzzy systemsA further report, entitled "Learning from data through the integration of qualitative models and fuzzy systems"  attempted to build on the earlier work from this research group and presents a methodology for the identification of the dynamics of non-linear patho-physiological systems once again by trying to learn from data. The key idea which underlies this approach consists once more of the integration of qualitative modelling methods with fuzzy logic systems. The major perceived advantage which derives from such an integrated framework lies in its capability both to represent the structural knowledge of the system at study and to exploit the available experimental data. As a result a functional approximation of the system dynamics can be determined and used as a predictor of the patient's future state. As a testing ground for their method the authors have considered the problem of identifying the response to insulin therapy in diabetes - using simulated data from AIDA to test out their approach.
Figure 1 summarises the scheme used for fuzzy system identification.
(Derived from Bellazzi et al ).
QSIM is the qualitative simulator that was developed by exploiting the physiological knowledge available in the literature, and in particular by referring to the studies presented in the main AIDA model paper  and the report of Berger & Rodbard . Based on this information Bellazzi and colleagues  initialized the membership functions of the fuzzy systems.
Therefore, this work  describes a novel approach to the identification of non-linear dynamic systems, which integrates fuzzy systems and qualitative models. The simulations obtained from a set of qualitative differential equations have been used to automatically encode the available knowledge in a fuzzy rule-based system; such a system is then tuned to a set of experimental data (from AIDA). The results obtained so far show that the presented framework generates fuzzy systems that may be used for a quick and reliable identification of non-linear systems in diabetes care .
Return to the Menu
How to improve fuzzy-neural system modeling by means of qualitative simulationIn further work by the same group the authors  have reported that the main problem in efficiently building robust fuzzy-neural models of non-linear systems lies in the difficulty to define a "meangingful" fuzzy rule-base. The authors' approach to the solution of such a problem is based on a hybrid method which integrates fuzzy systems with qualitative models. Bellazzi and colleagues introduce qualitative models to exploit the available, although incomplete, a priori physical knowledge of the system with the goal to infer, through qualitative simulations, all of its possible behaviours. The authors show that a rule base, which captures all of the distinctions in the system states, is automatically generated by encoding the knowledge of the system dynamics described by the outcomes of its qualitative simulation. Such a rule-base properly initializes a fuzzy identifier, which is then tuned to a set of experimental data .
Once again, simulated data coming from AIDA was used for training and testing, with data sampled from the AIDA program every 15 minutes.
Return to the Menu
Adaptive controllers for intelligent monitoringIn their report entitled “Adaptive controllers for intelligent monitoring”, Bellazzi and colleagues , have also described an approach based on the usual scheme of diabetes out-patient management, based on (i) a period evaluation of patients’ metabolic control performed by the physician, and (ii) patient-tailored tables for self-adjustment of insulin dosages. Following this scheme the authors have defined a system built on a two-level architecture. The High Level Module exploits both medical knowledge and clinical information in order to assess an insulin protocol, defined in terms of insulin timing, type, and total amount. The High Level Module exchanges information with the Low Level Module in order to define the control actions to be taken at the low level, as well as to periodically evaluate protocol adequacy on the basis of patient data. The goal of the Low Level Module, whose characteristics can be adaptively modified by the high level module, is to suggest the next insulin dosage depending on the actual blood glucose measurement and a certain pre-defined insulin delivery protocol. The Low Level Control Module is based on an adaptive controller, consisting of a fuzzy set controller and an ARX (Autoregressive eXogenous input) Model. The scheme presented by the authors  may be conveniently viewed in a telemedicine context, in which the low level controller is implemented on a portable device communicating to the high level controller, implemented on a remote computer.
The low level controller for this work  was tested using the AIDA diabetes simulation package. Bellazzi and colleagues simulated a patient weighing 40kg. They considered the following protocol: three injections per day of regular insulin, in correspondence to each meal, and two injections of NPH (intermediate-acting) insulin, at lunch time and at bed time. The total amount of daily insulin was 1 U/kg, and the proportion of the total amount for each dosage followed this scheme: Before Breakfast (7 am) 25% of regular insulin, Before Lunch (1 pm) 20% of regular insulin and 10% of NPH insulin, Before Dinner (6 pm) 20% of regular insulin, at Night Time (10 pm) 25% of NPH (intermediate-acting) insulin.
The authors compared the performance of four different control strategies for the implementation of the regulator.
The first strategy (a) just used a protocol without low level control (i.e. an open-loop strategy).
The second strategy (b) was an implementation via fuzzy rules of a decision rule set routinely used by patients in their self-monitoring activity; in other words it is a fuzzy controller having as an input variable the difference between the actual and the desired blood glucose level (i.e. a rule-based strategy).
The third strategy (c) was based on a fuzzy controller exploiting the ARX model predictions.
The fourth strategy (d) used the fuzzy controller operating with ‘perfect’ predictions - i.e. the blood glucose level that would be obtained if the open-loop strategy was followed.
The authors simulated the control system over 192 hours (8 days) with control actions and measurements at each meal (i.e. three times per day at 7 am, 1 pm and 6 pm). Bed time was taken to be 10 pm. Finally, it was assumed that the output measurements would be affected by random Gaussian noise.
It was observed that the rule-based control (strategy b) involved insulin adjustments of positive and negative signs, whereas the Fuzzy controller with the ARX model (strategy c), and the Fuzzy controller with ‘perfect’ predictions (strategy d) have only positive arguments. The authors concluded that these observations were related to the bad control performed with strategy (b) - which produced oscillations in control actions as well as in the patient’s blood glucose level. While initial prototype testing using data from AIDA simulations has been useful and encouraging, further testing of these approaches with real patient data are clearly required.
Return to the Menu
A distributed system for diabetic patient managementIn reference , entitled “A distributed system for diabetic patient management”, Bellazzi et al describe a telemedicine-based prototype for diabetes patient management. They present its architecture, the technical solutions adopted, and the methodologies on which it is based. The system is designed to provide decision support in a distributed environment, and is composed of two modules; (1) a Patient Unit and (2) a Medical Unit, connected by telecommunications services. The authors outline how the two modules can interact to perform effective monitoring and a cooperative control of glucose metabolism. In particular, Bellazzi and colleagues detail the data analysis tasks performed by the two units and how the results are used to assist patients and physicians in revising and adjusting the therapeutic protocol. The reported prototype implementation uses HTTP as the communications protocol and HTML pages as the graphical user interface.
The authors describe how the output of the reasoning module of the system is an ordered list of alternative diabetes regimens that should be able to solve the metabolic problems detected by the system. These alternative regimen protocols are presented to the physician who can then try them out using the AIDA simulator  and choose the most suitable one.
Return to the Menu
Protocol-based reasoning in diabetic patient managementIn reference , entitled “Protocol-based reasoning in diabetic patient management”, Montani and colleagues propose a system for teleconsultation in the management of patients with insulin-dependent diabetes mellitus (IDDM), accessible through the use of the Internet. The prototype is able to collect blood glucose monitoring data, analyse them through a set of tools, and suggest therapy adjustments in order to tackle the identified metabolic problems and fit these to the patient’s needs. The program tries to generate advice and employs it to modify the current therapeutic protocol, presenting the physician with a set of feasible solutions, from which he / she can choose the most appropriate one.
For this work, in order for the prototype to be able to calculate the effectiveness of a given food intake, food ‘activity’ was calculated using the AIDA model approach .
In all three Masters Theses [1-3], and subsequent publications from this group [4,5,8-11], AIDA was used as a simulator of blood glucose dynamics to test out the decision support prototypes.
Return to the Menu
Testing decision support prototypes #2:McCausland and colleagues from the University of Melbourne, Victoria, Australia have been developing an expert system to advise on insulin dosage adjustment in diabetes. The approach uses a combination of general rules, and rules which can be extracted from patient data - the idea being to produce a knowledge-based system which is able to 'learn' from the data and as a result automatically fine-tune the rules. These researchers are using simulated data from AIDA for initial testing of their decision-support prototype (McCausland, personal communication, 1999).
Return to the Menu
Testing decision support prototypes #3:Staite from University College Northampton in England has been developing a rule-based expert system to try and assist in patient insulin therapy self-management . The program uses IF... THEN... type production rules and concentrates on insulin-dosage adjustment for insulin-dependent (type 1) diabetic patients. The prototype is based around the user inputting blood glucose values taken at pre-determined times of the day - e.g. before each meal. Staite has made use of simulated blood glucose data from AIDA for initial testing of this decision-support approach.
Return to the Menu
Artificial Neural Networks - Background:An Artificial Neural Network (ANN) is an information processing model inspired by the way the densely interconnected parallel structure of the brain is thought to process information. In information technology terms an ANN is a computer system made up of a number of simple, highly interconnected processing elements which process information by their dynamic state responses to external inputs. ANNs are loosely modelled on the neuronal structure of the mamalian cerebral cortex, but obviously on a much smaller and less complex scale.
ANNs are believed to be of use in situations which involve the identification of highly non-linear and / or empirical systems; situations which conventional computer systems are less reliable at solving.
A simple example of an ANN making use of a three layer architecture is shown in Figure 2. These layers are constructed from a number of interconnected nodes which contain an activation function. Patterns are presented to the network via the input layer, which communicates to one or more hidden layers where the actual processing is done via a system of weighted connections. The hidden layers then link to an output layer (Figure 2) .
Figure 2. Multilayered artificial neural network. (Derived from Pender ).
Most ANNs contain some form of learning rule that modifies the weights of the connections according to the input patterns that are presented. Although there are many types of training algorithms used by neural networks, much of the work described in this section has made use of the ‘back-propagation’ training algorithm.
Using this technique input vectors and the corresponding output vectors are used to train a network until it can approximate a function, associating input vectors with specific output vectors. It is believed that networks with biases, a sigmoid layer and a linear output layer should be capable of approximating most functions with a finite number of discontinuities .
Apparently well trained back-propagation networks tend to give reasonable answers when presented with inputs that they have never seen before. Typically a new input will lead to an output similar to the correct output for input vectors that were used in training, and that are similar to the new input being present. This generalisation property makes the approach of interest to researchers in the diabetes field. As such, theoretically, it should be possible to train a neural network on a representative set of input / target pairs and get reasonable results without necessarily training the network on all possible input / output pairs.
Return to the Menu
Artificial Neural Networks - University of Strathclyde Prototypes:Dr. Bill Sandham’s group from the Institute of Communications and Signal Processing (Department of Electronic and Electrical Engineering) at the University of Strathclyde (Glasgow, Scotland, U.K.) have developed a number of ANN prototypes, and have done some interesting preliminary evaluation work. For the following research the MATLAB Neural Networks Toolbox software was used to create various ANN prototypes which were then trained using data from AIDA and / or clinical data. Using simulator generated data was described as removing “all the ethical and practical problems associated with collecting data from real patients” , and seems to have facilitated the rapid development / prototyping of a number of novel neural networks.
Return to the Menu
Artificial Neural Networks - University of Strathclyde - Prototype #1:The first ANN prototype, developed by Pender , under Dr. Sandham’s supervision made use of a hidden layer of 10 tan-sigmoid neurons to receive inputs directly and then broadcast their outputs to a layer of linear neurons that computed the network output (Figure 3). All weights and biases within this ANN were initialised with random values and the network was trained using training parameters and an error goal. The error was simply the difference between the neuron response and the desired or target vector. This network prototype was trained using an approximation of Newton’s method called the Levenberg-Marquardt technique; which is apparently more powerful and sophisticated than the more commonly applied alternative “gradient descent” method.
Figure 3. AIDA Artificial Neural Network Architecture. (Derived from Pender ).
This ANN was trained to predict blood glucose levels 2 hours ahead using the data from case scenario 0001 ("Joy Wilson") in the AIDA database. A prediction time of 2 hours was adopted because it was felt that a longer time could become more inaccurate due to meals. Input data used for this approach included (i) blood glucose and (ii) plasma insulin levels, as well as the AIDA glucose fluxes (iii) peripheral glucose uptake, (iv) net hepatic glucose balance, (v) carbohydrate absorption from the gut, and (vi) renal glucose excretion. The data were read off the AIDA graphs every 15 minutes. The data taken at 15 mins, 30 mins, and 45 mins past each hour were used as the training set. The remaining data was kept aside to be used as test data (previously unseen by the ANN). The main findings are shown in Table 1.
Trained with Tested with Average Error | AIDA – ANN | mmol/l (mg/dl) Relative Error | AIDA – ANN | / AIDA %
Patients 1 + 2
Table 1. Average and relative errors of an artificial neural network. (Derived from Pender ).
This ANN prototype was trained with a representative set of one patient’s diet and insulin schedule and compared to the simulated original values. It predicted blood glucose levels to within 3% of the actual blood glucose level. However when the neural network was trained on the data from one simulated subject (patient 1), less accurate predictions were made for another separate (different) subject (patient 2) . These results are shown graphically in Figure 4a and Figure 4b.
Figure 4a. Results from artificial neural network (ANN) trained and tested with data from same AIDA simulated 'patient'.
(Derived from Pender ).
Figure 4b. Results from artificial neural network (ANN) tested with data from different AIDA simulated 'patients'.
(Derived from Pender ).
It will be interesting to see how truly generalisable these ANN are, in practice, when trained with a large amount of data and tested against a new, different patient - separate from the training set.
Return to the Menu
Artificial Neural Networks - University of Strathclyde - Prototype #2 - Study #1:This research has been taken forward in the same laboratory by Sandham and colleagues  who have been investigating making clinical blood glucose predictions using more sophisticated artificial neural networks .
As highlighted above, ANNs are said to be particularly useful in situations which involve the identification of highly non-linear and / or empirical systems. They need to be trained on sets of patterns which display typical features of the system, but once trained, are meant to be generalisable using a variety of other data; the knowledge / experience acquired through training being embedded in the weight matrices of the ANN. Furthermore, through a dynamic learning process ANNs are meant to be able to assimilate information on a continuous basis.
For the second prototype a recurrent artificial neural network was adopted, since this has been reported to demonstrate superior performance for prediction problems with short term predictive accuracies ranging from 70-90% (in other fields) .
The recurrent ANN, as introduced by Elman (Figure 5) , has delays in the feedback loops at the outputs of the recurrent layer, which enable previous time-step values to be used in the current time step.
Figure 5. Architecture of Elman recurrent ANN  used for blood glucose prediction.
(Derived from Sandham and colleagues .)
Basically, the Elman recurrent network is a 2-layer network with feedback from the first layer output to the first layer input. This recurrent connection allows the Elman network to both detect and generate time-varying patterns.
Training for this recurrent ANN was performed using back-propagation incorporating a momentum term and an adaptive learning rate. Two separate activation functions were employed: neurons in the recurrent layer used a tan-sigmoidal function, whereas neurons in the output layer used a linear function. By inspection it was found that 95 recurrent layer neurons gave the best results.
For testing out this approach, initially six patients from the Diabetic Outpatient Department of Glasgow Royal Infirmary were selected. However due to the number of blood glucose measurements required, only the results from two patients could eventually be used.
Comparisons of the ANN with actual measured data for these two patients showed that most of the ANN predictions were very close to the measured values provided by the patients’ blood glucose meters (differences of 1.5 mmol/l [27 mg/dl] or less). However given that data were only available from two patients, a further study has been undertaken by Dr. Sandham using the AIDA diabetes simulator to provide a much larger amount of simulated patient data to more fully train and test the ANN.
Return to the Menu
Artificial Neural Networks - University of Strathclyde - Prototype #2 - Study #2:The aims and objectives of this further study  were to (i) harvest diabetic patient data using the AIDA diabetes simulation package, (ii) train the recurrent ANN with the generated data, and (iii) compare AIDA’s simulated blood glucose levels with predicted blood glucose levels from the ANN.
In total 50 data sets were harvested from the PC AIDA software (all from one case scenario), with 6 samples per data set. Each sample comprised an AIDA simulated blood glucose level, a carbohydrate intake (meal) in grams, and 2 doses of (short- and intermediate-acting) insulin, and a target predicted blood glucose level a set time after the sample.
Carbohydrate intake and insulin dose were changed for each data set. All other parameters were kept constant. The first 40 data sets were used to train the network and the final 10 data sets were used to test it.
The average error of this approach was reported as being 0.7 mmol/l [13 mg/dl] (a relative error of 11%). It is important to stress, however, that these data are based on only one AIDA case scenario, albeit with multiple simulations.
Nevertheless, with these test results the authors have been encouraged to expand the AIDA-based artificial neural network to include all the variable parameters included in the AIDA diabetes software simulation package.
The next step would then be to train this new network and test it to see if a similar standard of prediction can be obtained with a more comprehensive ANN.
As intimated above, it will also be interesting to see how well such an ANN trained with say data from 30 separate simulated AIDA patients - manages when presented with a completely new case.
Looking ahead, the interest in being able to predict blood glucose levels accurately is based on the fact that if such reliable and accurate BG predictions were possible, then therapy planning opportunities would clearly arise. With an ANN approach, the rationale is that a therapy optimiser could be encoded alongside the blood glucose predictor, within the ANN (Figure 6).
Figure 6 summarises the schematic of an ANN blood glucose predictor and ANN therapy optimiser. (Derived from Sandham and colleagues ).
The ANN predictor enables the predicted blood glucose level at time k+1 (BGLk+1) to be derived from the actual blood glucose at time k (BGLk), and the anticipated diet, exercise and insulin regimen at time k (DEIk). In addition previous values of these parameters, via the delay units (D), are used to optimise performance. All these parameters are then employed, together with the target BGL (BGLT) at the time k+1, to produce an optimum therapy plan in terms of diet, exercise and insulin regimen (DEIk+1). After suitable ANN training, weight matrices can be obtained for each patient, for either glycaemic prediction or attempted therapy optimisation.
One word of caution, however, should be sounded. This is to reinforce the fact that the AIDA model is a steady-state model of glucose-insulin interaction. As a result the model does not simulate the transient conditions that might result following a change in the insulin or dietary regimen. Rather the program simulates the longer-term effects of those changes, some 48-72 hours later. Given this, while the AIDA program is fine to be used for initial lab / bench-testing of various prototypes, it will be self-evident that there is still a very clear need eventually for proper clinical testing with real patient data.
In this respect, when used as described in this section of the Website, AIDA should theoretically be able to facilitate the rapid prototyping, development and initial testing of new / novel computational approaches; although real patient data for formal testing will still be required thereafter. However there is no reason that simulated data cannot be used to initially form the basis of (‘underpin’) the experience / knowledge encoded within an artificial neural network.
Return to the Menu
Artificial Neural Networks #2:Haque  from Brunel University in London, England has also used AIDA to provide blood glucose data to train an artificial neural network (ANN) of human carbohydrate metabolism in type 1 diabetes mellitus. Using a standard backpropagation network algorithm he developed an ANN to try and predict blood glucose levels for insulin-dependent diabetic patients following changes in either insulin therapy or carbohydrate intake. The network was trained using data collected from AIDA. The choice regarding the number of hidden nodes, and system parameters like the initial weight distribution, learning rate, etc were determined by investigating the network through the training process.
Input parameters for the ANN included the blood glucose level, carbohydrate intake, and insulin regimen at time, t. Patient specific parameters were not included. The output value was the predicted blood glucose level at some future time, t+1. Training the ANN required 150 data sets. To expedite this process 'AIDA on-line' was used as this allows easy access to the required data in an electronic form. Furthermore random values for the input carbohydrate range (0-80g) and insulin range (0-40 units) were used - generated with the RAND() function in Microsoft Excel - to permit as wide a variety of different cases to be simulated .
The ANN was batch trained with the computer left running for 2 days (approximately 40 hours) to try all the hidden nodes and other system parameters until the predictive error was reduced to about 10%. Each node was trained for up to 3,000 cycles before trying the next node. The best results were obtained using 23 hidden nodes .
Following training, the ANN was tested, first using 50 data sets that the neural network had seen before, and then using a further 50 data sets that the neural network had never seen before. The output result of each set was compared with the actual data obtained from the AIDA model. The ANN gave correct results (to the nearest integer) for 44 out of the 50 known data sets (12% error). For the 50 unknown data sets the ANN gave correct results for 38 cases (24% error).
Haque has done some very interesting work. As he highlights in his report  there are various ways that the predictive capabilities of the AIDA-trained ANN might be improved. He suggests using some knowledge or algorithm such as Jacob's Enhanced Back Propagation method to automatically adjust parameters such as initial weight and the momentum factor - rather than just setting these by inspection. Additional training could also be done - maybe up to 10,000 cycles for each hidden node - rather than just 3,000. Additional hidden layers might also be added. Furthermore it is important to note that not all components of the AIDA model were represented in the ANN. In particular, patient specific parameters for the kidneys (such as the renal threshold of glucose and renal function), insulin sensitivity (liver and peripheral) and the patient's weight were not included. It is to be expected that incorporating these clinical parameters in future work could lead to an improvement in the predictive accuracy of the ANN.
Return to the Menu
Artificial Neural Networks #3:Chang and colleagues  have developed an internet-based home monitoring prototype for diabetes care. Their approach proposes a tele-medical expert system to communicate the results of home monitoring of diabetes to a central hospital database, and to the physician in-charge of the patient. The expert system planned for decision support uses a neural network.
The strategy of the expert system uses a mapping method with back-propagation training of the neural network. With the back-propagation method, the weighting factors are controlled to reduce the performance function - that is the difference between the actual and desired network outputs. The basic back-propagation learning method adopted by Chang et al  aims to update the network weights and biases in the direction in which the performance function decreases most rapidly.
In this network, there are five inputs: (i) present blood glucose level, (ii) carbohydrate intake, and the amounts of (iii) short-, (iv) intermediate- and (v) long-acting insulin. The outputs from this network are the blood glucose levels that are predicted a few hours later. Figure 7 overviews the structure of the neural network.
Figure 7. Structure of the neural network model (derived from Chang et al ).
The function of this approach is to find out patients’ response patterns and try & estimate their blood glucose levels in the near future.
The system was evaluated by Chang and colleagues using test case scenario data obtained from the AIDA diabetes simulator .
Return to the Menu
Dietary assessmentYates and Fletcher  from the University of Liverpool in England have studied 3 published models of the gut to assess how well they were able to predict the appearance of glucose following the ingestion of a carbohydrate meal. They found the AIDA model to give the best results of the models tested  - although it is recognised that carbohydrate content only forms one component of an ordinary meal.
Yates and Fletcher  have also written in a more recent report that: “The glycaemic response of an insulin-treated diabetic patient goes through many transitory phases, leading to a steady state glycaemic profile following a change in either insulin regimen or diet. Most models attempting to model the glucose and insulin relationship try to model the effect of oral or injected glucose rather than that from the digestion of food. However, it is clear that a better understanding of the glycaemic response would arise from consideration of intestinal absorption from the gut. It is assumed that this type of absorption can be modelled by a so-called glucose appearance function (systemic appearance of glucose via glucose absorption from the gut) predicting the glucose load from the food. Much research has been carried out in the areas of hepatic balance, insulin absorption and insulin independent / dependent utilization. However, little is known about intestinal absorption patterns or their corresponding glucose appearance profiles.
The strategy under investigation herein is to use deconvolution or backward engineering. By starting with specific results i.e. blood glucose and insulin therapy, it is possible to work backwards to predict the glucose forcing functions responsible for the outcome. Assuming compartmental consistency, this will allow a clearer insight into the true gut absorption process. If successful, the same strategy can be applied to more recent glucose and insulin models to further our understanding of the food to blood glucose problem” .
The authors investigated the AIDA model of glucose and insulin interaction. This model simulates the steady state glycaemic and plasma insulin responses, independent of the initial values from which the simulation is started. Glucose enters the model via both intestinal absorption and hepatic glucose production. Yates and Fletcher considered a 70 kg male insulin-dependent diabetic patient with corresponding hepatic and insulin sensitivity parameters of 0.6 and 0.3 respectively. Net hepatic glucose balance was modelled piecewise by linear and symmetric functions. A first-order Euler method with a step size of 15 minutes was employed. For the simulation, only Actrapid and NPH injections were considered. The injection of insulin and the glucose flux from the gut were started simultaneously to avoid any delay associated with gastric emptying.
The systemic appearance of glucose was compared from two view points, not only to assess the strategic principle, but also to assess the suitability of the AIDA model. The first is a forward prediction using the compartmental structure. This analysis involves the rate of gastric emptying without time delay. The second is a backward prediction from experimentally observed blood glucose profiles. Investigations involved porridge, white rice and banana containing the same carbohydrate content (25 g). Results obtained from the first analysis were dependent on the rate of gastric emptying, especially its ascending and descending branches. Results from the second analysis were dependent on the dose and type of insulin administered. Both predicted profiles showed consistency with physiological reasoning, although it became apparent that such solutions could be unstable. Furthermore, both types of prediction were similar in structure and appearance, especially in simulations for porridge and banana. This emphasized the consistency and suitability of both analyses when investigating the compartmental accuracy and limitations within a model.
The new strategic approach was deemed a success, and the AIDA model was found to be "appropriate" . Yates and Fletcher suggested that a gastric emptying curve with a possible gastric delay is the way forward in regulating the appearance of glucose via gut absorption. The AIDA model gastric curve is described by either a trapezoidal or triangular function dependent on the carbohydrate content of the meal. However, it was apparent to Yates and Fletcher from their results that carbohydrate content is only one factor in carbohydrate absorption, and that to further improve realism further progress must inevitably involve other food characteristics and properties .
Return to the Menu
Diabetes model - further developments #1:Cobelli and colleagues  from the University of Padova in Italy are in the process of developing a new physiological model of glucose-insulin interaction in type 1 diabetes mellitus. This is intended to encompass newly acquired physiological knowledge about the time course of endogenous glucose production during a meal , and about the effect of glucose and insulin signalling on glucose utilization . It is also meant to apply more accurate descriptions of the action profiles of insulin following subcutaneous injection . As part of their testing procedure the researchers have been comparing the performance of their new model with other models of type 1 diabetes mellitus, including AIDA [6,24].
Furthermore the new Cobelli model is currently being extended to describe a Type 1 diabetic patient in order to provide a test bed for examining various data analysis techniques and control strategies , in much the same way that other researchers highlighted in this section of the Website have been making use of AIDA.
Return to the Menu
Diabetes model - further developments #2:Butler , Strachan and colleagues from the Robert Gordon University in Aberdeen, Scotland, U.K. have been working on a 'Lifestyle Model' for insulin dependent diabetic patients. AIDA is one of several models they are reviewing; the eventual aim being to try and improve the control and quality of life of diabetic patients by developing an enhanced model of the processes that are known to affect diabetes mellitus.
Return to the Menu
Diabetes model - further developments #3:Escreet  from Staffordshire University, England has also reviewed the AIDA program, as part of work to develop a more comprehensive diabetes model. As well as critiquing the application, he wrote:Escreet attempted to enhance the AIDA simulation approach by the addition of glycaemic indices for foods and by taking exercise into account - the aim being to provide a more comprehensive model of glucose-insulin interaction in diabetes. Such enhancements to the model are of some importance, but in practice they are not without their difficulties. For instance, the glycaemic index is a parameter that works quite well for single foods. However when one tries to combine foods into meals the process becomes much more complex, and the parameter less useful in practice. Similarly with exercise - quantifying the level of activity can be quite problematical - unless one goes for a relatively simple qualitative representation with low, medium and high levels of exertion. Further research to address these issues is clearly required.
"AIDA is an extremely useful blood glucose simulation tool that contains a wealth of information for the user. The large number of example cases with full descriptions should be enough to get the user familiar with how to use the program as well as imparting further, more practical knowledge of how the glucose-insulin reaction works within the body."
Return to the Menu
Web based diabetes simulator: AIDA on-line
One of the most successful research projects to-date involving AIDA has been the development of a Web-based interactive educational diabetes simulator. This resulted from an approach by two students at North Carolina State University (NCSU), North Carolina, U.S.A. who had been set as a senior bioengineering design project to create a simulator of glucose-insulin interaction on the Web. After rather more effort than normally goes into such a project 'AIDA on-line' was produced , the system first going live on the Web in December 1997.
This can be accessed via this link and permits AIDA's interactive diabetes simulations to be run from anywhere in the world - from any computer platform (Mac, PC, Linux, UNIX server, etc) - provided it is connected to the Internet and has a graphical display. Users have even reported feedback using 'AIDA on-line' via Web TV.
'AIDA on-line' offers a familiar mouse / Windows graphical user interface - all interactions taking place via a standard Web browser window. The service is fast to use. Running a simulation from a computer in Seattle, Washington State, U.S.A. (with the server residing in London, England, U.K.) usually only takes 1-2 seconds. Further information about this development can be found by clicking here. Figure 8 summarises the basic structure of the 'AIDA on-line' system.
Figure 8. Basic structure of 'AIDA on-line'. HTML = HyperText Markup Language. The ‘AIDA on-line’ homepage (http://www.2aida.net) presents the user with various simulator options and case scenarios. A series of dedicated (Common Gateway Interface) CGI-BIN scripts written in Perl v5.0 are used to read case scenario data and insulin and carbohydrate profiles from the database (DB). The plasma insulin and blood glucose profiles are computed using a further Perl script which contains the AIDA model differential equations (and which makes use of some temporary storage space on the ‘AIDA on-line’ server). Output from the simulator is returned to the user in HTML format for display by a Web browser. Further details about how ‘AIDA on-line’ works can be found elsewhere at this Website by clicking here.
Since logging of the number of simulations was started in August 1998 - over 155,000 interactive diabetes simulations have been run at 'AIDA on-line' .
Return to the Menu
On-going research using AIDA
Various other research projects involving AIDA are currently underway, although these have not yet reported their results. Nevertheless from its release to other researchers in 1992 / 93 - and on the Internet in 1996 - it is interesting to see how quickly such a simulation program can be adopted for wider research application. This should encourage more researchers to consider making use of the Internet for the distribution of their work. As in the case of AIDA - such distribution may not only possibly be of use to patients directly - but may also actually help to promote further research. In addition, publication of all AIDA model and systems details - as can be found both in the literature [6,33] and on the Internet - may help to support further research work in this field (see below).
However, it is very possible that the ventures reported in this Research Section of the Website under-represent the totality of research projects that have actually been making use of the AIDA software. This may be because since this Web page first was made live on the Internet, the information has been openly and freely available. As a result less students and researchers have actually needed to get in contact prior to embarking on research projects making use of AIDA.
Return to the Menu
Further ideas - generating simulated blood glucose data - for training / validation
The computer-science literature is full of descriptions of decision-support prototypes which attempt to provide therapeutic advice for patients with diabetes. Such prototypes use a wide variety of different computational techniques (see refs [34-38] for an overview). While academically and intellectually such approaches may be of interest in their own right - often the prototypes stumble at the first real medical hurdle - that of testing or validation. Computer scientists with access to computer labs and computer facilities may have limited access to patients with diabetes. Therefore achieving clinical collaboration to take forward the verification, validation, and clinical evaluation of such prototypes is not without its difficulties.
AIDA cannot offer a complete solution to this problem. However it does offer a way for computer scientists and medical informaticians to generate some reasonably realistic blood glucose data - from a wide variety of 'virtual diabetic patients'. Such data can then be used for testing other computer-based prototypes. In this respect testing of a decision-support prototype against data from AIDA's 40 example case scenarios should at least help to identify areas where further refinements or work are required. Also while such simulator-based testing cannot replace clinical testing with real patient data, it may be that positive simulator-based test results can help to encourage clinical collaboration to take forward the validation testing with real patient data.
In support of this - as for patient, relative, carer or health-care professional use - AIDA is freely available via the Web. Click here to download a copy. If you require access to a large quantity of blood glucose data, for example to train a neural network, you may find the 'AIDA on-line' diabetes simulator of more use than the PC version.
'AIDA on-line' has a feature which allows the raw simulated blood glucose data to be outputted in electronic form. To make use of this facility run a simulation and then click on the underlined html Data link in the top right corner above the blood glucose graph. This will provide access to the 'raw' simulation data at 15 minute intervals.
How you make use of these data is up to you. However you could select certain data points from this (e.g. before meals) and feed these through your prototype and also provide them to some clinicians for their assessment of the advice from your computer program. Alternatively you could make use of all the data passing it to a neural network - to provide some preliminary training for that network. As highlighted above, it should be stressed that such data cannot replace similar testing / training with real patient data. However it is hoped that the ready availability of simulated blood glucose data may save a considerable amount of time - especially during the early stages of a new prototype's development.
Return to the Menu
Learned space of parameter interactions using simulated blood glucose dataA small group of researchers housed at the NASA Independent Verification and Validation Facility (a Software Research Lab which is now a division of Goddard Space Flight Center) have been engaged in utilizing machine learning as an analysis technique, the goal of which is to provide importance hierarchies of parameters in determining changes in the model output.
The background to this work is the fact that recent high profile satellite losses have highlighted NASA’s need for quality software. The NASA Software Research Laboratory conducts applied research into advanced software analysis technologies, through tool development, case studies and pilot projects.
The technique overviewed below, originally developed for software risk assessment is, however, not limited to that domain. By incorporating other data / another model the researchers expect to be able to use the method in other areas. For instance, they are currently applying the approach to diet and insulin dosage adjustment planning for patients with diabetes. Using ‘AIDA on-line’ to provide blood glucose data, they hope to learn the simplest actions that individuals with diabetes should be able to take to maintain their blood glucose at the correct level.
The researchers have presented their original technique with respect to a software project risk model , and are now seeking to extend their research into the diabetes domain, utilizing the AIDA model.
Their original approach has been using mass simulation data in a Monte-Carlo style analysis of model behaviors, which are learned by a standard Machine Learner technology. They then explore this learned space of parameter interactions to find importance hierarchies of parameters with respect to influencing changes in the output values.
In the original software project risk example this concerned such things as the influence of programmer experience on project risk. In terms of AIDA, or a diabetes model in general, it is hoped this may include such things as the influence of calorific intake at certain times of the day on levels / frequency of a type of insulin injection, or some other index measure. It is expected that what could result from a trusted model or data would be a general ranking of importance for the factors involved, aiding in resource / effort allocation in controlling or monitoring those factors.
Clearly a large amount of data is required for any information which is actually usable in a patient's regimen. In this respect the researchers are currently using ‘AIDA on-line’ to generate large quantities of simulated blood glucose data, and require more than 60,000 simulations to provide sufficient entries in their dataset to perform a stable machine learner summarization with which to undertake their further analyses.
Three main steps are involved in the process, when applied to diabetes:[A] Collect blood glucose data or use a model to provide simulated valuesIn the diabetes domain it is hoped that [C] will permit actions to be identified that may improve glycaemic control.
[B] Perform machine learning to convert the examples into an ensemble of decision trees
[C] Use dedicated software to find mitigation strategies that change classifications (risk) in a majority of the decision trees .
However in the future the researchers hope to present a technique which may be utilized in the presence of lesser amounts of data, possibly allowing a patient to build their own personal analysis model showing for that particular individual which factors seem to most strongly influence his / her symptoms.
Further information about this approach can be obtained from Erik Sinsel, NASA / WVU Software Research Laboratory, 100 University Drive, Fairmont, WV 26554, USA
Return to the Menu
Performance monitoring of closed-loop insulin delivery devicesOwens and Doyle  have been addressing a different research question. They have described using the Bergman minimal model  and the AIDA model  as virtual diabetic patients in their research on performance monitoring of closed-loop insulin delivery devices.
With closed-loop insulin delivery devices, as with any automated device, malfunction is possible. Hence the research issue that Owens & Doyle have been targeting is the fact that a measure of performance would be helpful to detect device failures that could violate hypoglycaemic and hyperglycaemic bounds of the patient. Therefore performance measures for the controller are an essential component of a closed-loop algorithm, to ensure safe operation . Intra-patient variability and faults due to insulin aggregation in the pump, sensor fouling, etc are all conditions that can occur and affect pump controller performance.
For this work, the researchers have studied the Bergman and AIDA models in the standard Internal Model Control framework, using a first order model for controller synthesis. They have tried to assess the effects of things like insulin aggregation and sensor fouling on controller performance .
While an Internal Model Control approach can provide adequate control of blood glucose levels using such a device, in a separate report the same authors have also reported studying the use of a Model Predictive Control approach . This technique apparently has an improved ability to handle the hypoglycaemic, hyperglycaemic and insulin dosage constraints of such systems. Furthermore the use of Model Predictive Control (MPC) in a biological context has previously proved successful with applications in the field of anaesthetics. As a result of using a MPC approach it is reported that tighter control can be obtained, limiting the effects of meals and other disturbances on glucose homeostasis .
In a separate, but related, piece of work from the same group - Parker and colleagues  have developed a model-based predictive control algorithm for blood glucose control in type 1 diabetic patients, using a closed-loop insulin infusion pump. Using compartmental modelling techniques a fundamental model of a patient with diabetes was constructed. The resulting nineteenth-order non-linear pharmacokinetic-pharmacodynamic representation is used in controller synthesis. The mathematical representation of the meal (carbohydrate) sub-model used for this work, was that developed for AIDA .
Return to the Menu
A business plan utilising telehealth-care technology within the homeConmy  has described a detailed business plan for utilising telehealth-care technology in the home. This is based on the premise of a diabetes nurse care manager in association with a primary care physician (general practitioner [G.P.]), and with the assistance of AIDA, being able to manage patients’ blood glucose levels.
This Thesis (from 1999) for its AIDA work focuses exclusively on a single AIDA paper from 1994  which cites earlier validation work of a predecessor of the AIDA knowledge-based system (KBS). Unfortunately the Thesis overlooks the fact that the AIDA model itself (separate from the KBS) is not accurate enough for individual patient glycaemic prediction or therapy planning. This is made clear at the AIDA Website (at: http://www.2aida.org/caveats) and in the AIDA software as well as in a number of research publications from 1996 and 1998 [46,47]. It is for this reason that AIDA has only been made available for educational / demonstration / teaching and/or self-learning purposes.
Connected with this, many researchers have hoped to be able to one day develop a computer system that might be able to assist in generating insulin-dosage adjustment advice. Indeed the literature [34-37] and this Web page are full of descriptions of prototypes attempting to achieve this. However, to date, producing a computer program which can generate reliable, accurate medical advice - and which can explain its reasoning - has not proved so easy in the diabetes field. As such the financial planning described in the Thesis  might be considered slightly premature. Nevertheless it is highlighted here as another example of a research project that has reported heavy reliance on the AIDA software.
Return to the Menu
Evaluation / validation research usage of AIDAAIDA, being widely available and completely free, could serve as a useful ‘test bed’ for generally establishing how best to evaluate educational medical software programs. Although many prototype medical applications have been described - the majority do not seem to be used so extensively so such evaluation issues have not been considered in great detail, especially in diabetes care. In this respect, to date, AIDA has generated some useful discussion about the best research methods to apply to properly validate and evaluate such programs [48,49], and it is expected that AIDA will stimulate further contributions to this debate in the future. Connected with this, because AIDA is so widely used it should be possible to start to address such evaluation issues in greater detail and potentially with a larger number of evaluators.
Figure 9 demonstrates a pyramid of different levels of evidence for formally evaluating educational / clinical use of a piece of software, that can be applied to a program like AIDA.
Figure 9. Highlights the different levels of evidence for formally evaluating educational / clinical use of a program like AIDA (derived from Lehmann ).
Level 1 studies (randomised controlled trials, RCTs) are clearly the ‘gold standard’ method for rigorously assessing such educational / clinical utility. Nevertheless useful information can also be obtained from less formal studies, and in many cases these can be easier to undertake and involve more subjects than RCTs. In the case of AIDA, under Level 5 (methodological verification and validation studies), a quantitative assessment was reported in 1994 to document the accuracy of the blood glucose simulations in a cohort of 30 patients with diabetes . While the simulations were shown to be unsuitable for individual patient glycaemic prediction and therapy planning, they have found widespread use for educational / demonstration / self-learning purposes where individual predictive accuracy is less critical [46,47].
Using the evaluation scheme shown in Figure 9, recent reports  would fit under Level 4 (anecdotal evidence including user comments and reviews). While such feedback has been very encouraging, the next stage in the evaluation process is clearly to undertake Level 3 observational studies (including the use of surveys and more formal questionnaires ) and prospective RCTs.
For RCT use, a standardised protocol for the evaluation of such diabetes simulation programs has been developed [52,53] and early pilot study (preliminary) results from a small number of patients (n=24) have shown this to be a viable method for formally evaluating programs like AIDA . (Further information can be found at this Website at: http://www.2aida.org/evaluate). However a larger number of subjects in multiple centers are clearly required to properly test out the use of this approach as a teaching tool.
Nevertheless, a wide range of users in many different parts of the world do seem to have identified AIDA as an accessible source of information about glucose-insulin interaction, and a useful supporting method for diabetes education. Taken together with the growing number of reports of user experience [55-60] collectively these comments highlight the potential for empowerment that some people feel can result from use of the program. Given this, it is suggested that the experience with this approach is sufficiently encouraging to warrant more formal, randomised controlled clinical studies to identify the actual clinical / educational role for such interactive diabetes simulators .
Return to the Menu
Other research uses of AIDA / Placing technical information on the WebPlacing the AIDA Technical Guide on the Web (at: http://www.2aida.org/technical) has yielded interesting and surprising results. This technical overview of the AIDA model, and how the simulator works, was first uploaded to the Internet in July 1998. In the 4 years from then until July 2002, the Web page has received over 7,000 visits. This illustrates how putting such material on the Internet allows a lot more people access to such technical / research information than would have been the case just through standard paper journals.
In addition, since AIDA moved to its own dedicated Website in October 2000 various more recent research articles have been made available on the Internet. These research papers can be accessed completely free-of-charge as portable document format (PDF) files. In the approximately 1.75 years from October 2000 to July 2002 there have been over 8,500 PDF file downloads from the AIDA Website. Even leaving out leaflets, questionnaires and consent forms which are also available as PDF files, there have still been over 6,200 downloads of research articles from the site. There are few researchers who could have serviced this sort of level of interest / reprint requests, manually, via post / Air Mail in previous years. It is expected that such articles (and the on-line AIDA Technical Guide) are two of the ways that information about AIDA has been disseminated so widely.
Once again, this illustrates how placing such material on the Web allows much larger numbers of people access to such research information than would have been the case just through regular libraries and standard hard copy reprint requests.
The conclusion from all this experience is that more researchers should consider making use of the Web to disseminate their programs / work. Not only can such Web-based distribution help patients and their relatives - in the case of AIDA through education - but it can also help promote and support further on-going research. In this respect, clearly AIDA has seen use - not only by patients, their relatives, students and health-carers - but also as a tool for research application in other areas of diabetes IT.
Return to the Menu
Have a research project of your own?
We maintain this 'Research Use' section as a way of highlighting one of the benefits of distributing work, like AIDA, via the Internet. Also we hope that more students / researchers embarking on their own diabetes-computing research projects may consider whether AIDA could be of use to them.
If you are thinking of undertaking a research project - and are wondering whether AIDA might be able to help - please feel free to contact the AIDA authors via the on-line AIDA contact form. We are happy to support such research use of the software - in whatever small ways we can - even from afar.
If you have already made use of AIDA for research purposes - please do get in touch. We are always interested to hear about a wider range of people's experience with the software.
Return to the Menu
Research publications on-line
If you are interested in research into the application of computers in clinical diabetes care - you may also like to check out the links below:
- Two part special journal issue entitled Application of information technology in clinical diabetes care. Part 1. Databases, Algorithms and Decision Support and Part 2. Models and Education.
- Review article entitled Application of computers in clinical diabetes care
- AIDA Technical Guide
- AIDA Model Graphics (with links to further graphs)
- AIDA Model Graphics (more detailed / printable summary)
- Available insulin preparations
- Diabetes-IT articles from Diabetes Technology & Therapeutics
- Other on-line research publications
Return to the Menu
References Guglielmann R. Non linear systems identification by using qualitative models and fuzzy systems. MSc Dissertation, University of Pavia, Italy, 1996 [in Italian].
 Bossi G. Neuro-fuzzy system identification based on qualitative models: an application to physiological systems. MSc Dissertation, University of Pavia, Italy, 1996 [in Italian].
 Telecco P. Metabolic control in diabetic patients: decision support systems. MSc Dissertation, University of Pavia, Italy, 1994 [in Italian].
 Bellazzi R, Ironi L, Guglielmann R, Stefanelli M. Qualitative models and fuzzy systems: an integrated approach for learning from data. Artificial Intelligence in Medicine 1998; 14: 5-28.
 Bellazzi R, Ironi L, Guglielmann, Stefanelli M. Learning from data through the integration of qualitative models and fuzzy systems. In: Lecture Notes in Artificial Intelligence 1997; 1211: 501-512.
 Lehmann ED, Deutsch T. A physiological model of glucose-insulin interaction in type I diabetes mellitus.
Journal of Biomedical Engineering 1992; 14: 235-242.
 Berger M, Rodbard D. Computer simulation of plasma insulin and glucose dynamics after subcutaneous insulin injection. Diabetes Care 1989; 12: 725-736.
 Bellazzi R, Guglielmann R, Ironi L. How to improve fuzzy-neural system modeling by means of qualitative simulation. IEEE t-Neural Networks 2000; 11: 249-253.
 Bellazzi R, Siviero C, Stefanelli M, De Nicolao G. Adaptive controllers for intelligent monitoring. Artificial Intelligence in Medicine 1995; 7: 515-540.
 Bellazzi R, Riva A, Larizza C, Fiocchi S, Stefanelli M. A distributed system for diabetic patient management. Computer Methods and Programs in Biomedicine 1998; 56: 93-107.
 Montani S, Bellazzi R, Larizza C, Riva A, d'Annunzio G, Fiocchi S, Lorini R, Stefanelli M. Protocol-based reasoning in diabetic patient management. International Journal of Medical Informatics 1999; 53: 61-77.
 Staite R. The use of IT-mediated patient self management in insulin therapy management. BSc Dissertation. University College Northampton, Northampton, UK, 1999.
 Pender JE. Modelling of blood glucose levels using artificial neural networks. Dissertation. University of Strathclyde, Scotland, 1997.
 Sandham WA, Nikoletou D, Hamilton DJ, Patterson K, Japp A, MacGregor C. Blood glucose prediction for diabetes therapy using a recurrent artificial neural network. In: Proceedings, EUSIPCO-98, IX European Signal Processing Conference, Rhodes Island, Greece, 1998; Vol. 11: pp. 673-676.
 Sandham, W.A., D.J. Hamilton, A. Japp and K. Patterson, (1998), Neural network and neuro-fuzzy systems for improving diabetes therapy. In: Proceedings, 20th International Conference of the IEEE Engineering in Medicine and Biology Society, 29 October - 1 November 1998, Hong Kong Convention and Exhibition Centre, Hong Kong, vol. 20: part 3/6, pp. 1438-1441.
 Patterson WD. Artificial Neural Networks - Theory and Applications. Prentice Hall, Singapore, 1996.
 Elman JL. Distributed representations, Simple recurrent networks and grammatical structure. Machine Learning 1991; 7: 195-225.
 Sandham WA. University of Strathclyde, Scotland, UK. Personal Communication, 2002.
 Haque A. Modelling human metabolism using neural network. Dissertation. Brunel University, London, UK, 1999.
 Chang SC, Ryoo SM, Yu SY, Ahn BH. Development of an internet-based home monitoring system for diabetes mellitus. In: Proceedings, 4th IEEE International Symposium on Consumer Electronics (ISCE’99), Melaka, Malaysia, 1999.
 Yates T, Fletcher LR. Prediction of a glucose appearance function from foods as investigated in the Lehmann/Deutsch model of glucose insulin interaction. In: Proceedings, Computers in Diabetes'96, 1996; Graz, Austria, p.29 (abstract).
 Yates TL. Forcing function for blood glucose - insulin models. MSc Dissertation. University of Liverpool, UK, 1995.
 Yates TL, Fletcher LR. Prediction of a glucose appearance function from foods using deconvolution. IMA Journal of Mathematics Applied in Medicine and Biology 2000; 17: 169-184.
 Cobelli C, Nucci G, Del Prato S. A physiological simulation model of the glucose-insulin system in type 1 diabetes. Diabetes Nutrition and Metabolism 1998; 11: 78 (abstract).
 Taylor R, Magnusson I, Rothman DL, Cline GW, Caumo A, Cobelli C, Shulman GI. Direct assessment of liver glycogen storage by 13C nuclear magnetic resonance spectroscopy and regulation of glucose homeostasis after a mixed meal in normal subjects. Journal of Clinical Investigation 1996; 97: 126-132.
 Cobelli C, Caumo A. Using what is accessible to measure that which is not: necessity of model of system. Metabolism 1998; 47: 1009-1035.
 Torlone E, Pampanelli S, Lalli C, Del Sindaco P, Di Vincenzo A, Rambotti AM, Modarelli F, Epifano L, Kassi G, Perriello G, Brunetti P, Bolli G. Effects of the short-acting insulin analog [Lys(B28),Pro(B29)] on postprandial blood glucose control in IDDM. Diabetes Care 1996; 19: 945-952.
 Cobelli C, Nucci G, Del Prato S. A physiological simulation model of the glucose-insulin system. In: Proceedings, First Joint IEEE BMES/EMBS Conference Serving Humanity, Advancing Technology 1999; Vol. 1: p. 999.
 Butler RA. Mathematical Modelling, Simulation and Forecasting of Insulin Dependent Diabetes. In: Proceedings, 3rd International Conference Saterra, Mittweida, Germany, 12 - 15 November 1997, pp. 81-88.
 Escreet A. Dietary Aid for Diabetics. Dissertation. Staffordshire University, England, 1999.
 Lehmann ED, DeWolf DK, Novotny CA, Gotwals Jr RR, Rohrbach RP, Blanchard SM. Dynamic interactive educational diabetes simulations using the World Wide Web. Diabetes Technology and Therapeutics 2001; 3: A17-A20 (abstract).
 Lehmann ED. Usage of a diabetes simulation system for education via the Internet. International Journal of Medical Informatics 2002; (in press) (letter).
 Lehmann ED, Deutsch T. AIDA2 : A Mk. II Automated Insulin Dosage Advisor. Journal of Biomedical Engineering 1993; 15: 201-211.
 Lehmann ED, Deutsch T. Application of computers in diabetes care - a review. I. Computers for data collection and interpretation. Medical Informatics 1995; 20: 281-302.
 Lehmann ED, Deutsch T. Application of computers in diabetes care - a review. II. Computers for decision support and education. Medical Informatics 1995; 20: 303-329.
 Lehmann ED. (Ed.) Special Issue: Application of information technology in clinical diabetes care. Part 1. Databases, algorithms and decision support. Medical Informatics 1996; 21: 255-378.
 Lehmann ED. (Ed.) Special Issue: Application of information technology in clinical diabetes care. Part 2. Models and education. Medical Informatics 1997; 22: 1-120.
 Lehmann ED. Application of computers in clinical diabetes care. Diabetes Nutrition and Metabolism 1997; 10: 45-59.
 Menzies T, Sinsel E. Practical large scale what-if queries: case studies with software risk assessment. In: Proceedings, 15th IEEE International Conference on Automated Software Engineering, (September 11-15, 2000), pp. 165-173, IMAG Grenoble, France.
 Owens CL, Doyle FJ III. Performance monitoring of closed-loop insulin delivery devices. Diabetes Technology and Therapeutics 2002; 4: 228 (abstract).
 Bergman RN. Toward physiological understanding of glucose tolerance. Minimal-model approach. Diabetes 1989; 38: 1512-1527.
 Doyle FJ III, Owens CL. Model based performance monitoring of diabetic patient systems. In: Proceedings, American Institute of Chemical Engineers, Annual Meeting, 2001 (abstract).
 Parker RS, Doyle FJ III, Peppas NA. A model-based algorithm for blood glucose control in type I diabetic patients. IEEE Transactions on Biomedical Engineering 1999; 46: 148-157.
 Conmy P. A business plan utilizing telehealth technology within the home. MSc Thesis. San Francisco State University, San Francisco, California, 1999.
 Lehmann ED, Hermanyi I, Deutsch T. Retrospective validation of a physiological model of glucose-insulin interaction in type 1 diabetes mellitus. Medical Engineering and Physics 1994; 16: 193-202 [Published erratum appears in Med Eng Phys 16: 351-352].
 Lehmann ED, Deutsch T. Computer assisted diabetes care: a 6 year retrospective. Computer Methods and Programs in Biomedicine 1996; 50: 209-230.
 Lehmann ED, Deutsch T. Compartmental models for glycaemic prediction and decision-support in clinical diabetes care: promise and reality. Computer Methods and Programs in Biomedicine 1998; 56: 193-204.
 Biermann E. Comment on Lehmann and Tatti: proposed controlled trial on simulators in diabetes education. Diabetes Technology and Therapeutics 2002; 4: 255-257 (letter).
 Lehmann ED, Tatti T. Randomized controlled trial design for simulator use in diabetes education: some issues for consideration. Diabetes Technology and Therapeutics 2002; 4: 258-269 (letter).
 Lehmann ED: The freeware AIDA interactive educational diabetes simulator - http://www.2aida.org - (1) A download survey for AIDA v4.0. Medical Science Monitor 2001; 7: 504-515.
 Lehmann ED. Further user comments regarding usage of an interactive educational diabetes simulator (AIDA). Diabetes Technology and Therapeutics 2002; 4: 121-135.
 Tatti P, Lehmann ED: A randomised-controlled clinical trial methodology for evaluating the teaching utility of interactive educational diabetes simulators. Diabetes Nutrition and Metabolism 2001; 14: 1-17.
 Lehmann ED, Tatti P: Questionnaires for a randomized controlled trial methodology to evaluate the teaching utility of diabetes simulation programs. Diabetes Technology and Therapeutics 2001; 3: 293-305.
 Tatti P, Lehmann ED: Preliminary results from a randomised controlled clinical trial for evaluating the teaching utility of an interactive educational diabetes simulator (AIDA). Diabetes 2001; 50(Suppl.2): A25 (abstract).
 Lehmann ED: Preliminary experience with the Internet release of AIDA - an interactive educational diabetes simulator. Computer Methods and Programs in Biomedicine 1998; 56: 109-132.
 Lehmann ED: Spontaneous comments from users of the AIDA interactive educational diabetes simulator. Diabetes Educator 2000; 26: 633-643.
 Lehmann ED: User experience with the AIDA interactive educational virtual diabetes patient simulator. Diabetes Technology and Therapeutics 2000; 2: 165-171.
 Lehmann ED: Short user comments (‘sound bites’) regarding usage of AIDA v4 - http://www.2aida.org - an interactive educational diabetes simulator. Diabetes Technology and Therapeutics 2000; 2: 663-666.
 Wilson DM: Diabetes simulators: ready for prime time? Diabetes Technology and Therapeutics 1999: 1: 55-56.
 Lehmann ED: The freeware AIDA interactive educational diabetes simulator - http://www.2aida.org - (2) Simulating glycosylated haemoglobin (HbA1c) levels in AIDA v4.3. Medical Science Monitor 2001; 7: 516-525.
Return to Top of Page If you like AIDA, why not display our logo on your home page? For more information about linking to the AIDA Website please click here.
Return to AIDA Website Home Page AIDA is a freeware diabetes software simulator program of glucose-insulin action + insulin dose & diet adjustment in diabetes mellitus. It is intended purely for education, self-learning and / or teaching use. It is not meant for individual blood glucose prediction or therapy planning. Caveats
This Web page was last updated on 19th October, 2003. (c) www.2aida.org, 2000-2003. All rights reserved. Disclaimer. For the AIDA US Mirror Site, please click here. For the Diabetes / Insulin Tutorial, please click here.