The aim of the proposed project is to develop and implement a web-based Medical Information Repository (MIR) integrated with a computerized medical Decision Support System (DSS) to promote advanced, quality healthcare information and services for all people, especially the rural communities, through the use of Information and Communication Technology (ICT) as well as Artificial Intelligence (AI) methodology.

 

The main objectives of the proposed project are as follows.

 

Outputs of the proposed project include

Novelty of the proposed project lies in the inclusion of an AI-based DSS into the MIR.  The ART family of neural networks will be utilized to design the DSS for early prognosis and diagnosis of stroke patients.  Unlike other intelligent systems existed in the literature, innovation of the proposed system lies in the autonomous learning behaviors of ART-based neural networks.  Once developed, the system is able to learn incrementally in real-time, non-stationary environments with minimum intervention from the system designer (computing or AI experts).  With this autonomous learning capability, domain users (medical practitioners) are able to train and fine-tune the decision support system, and to assume “ownership” of the system, hence overcoming resistance by non-computing/AI users to apply such a computational tool in their work.

 

 

A stroke is an interruption of the blood supply to any part of the brain, resulting in damaged brain tissue.  It is one of the most common neurological problems and a leading cause of mortality in the world.  Stroke is of increasing importance owing to a rapidly aging population as well as from the multiplicative effects of both the economic and psycho-social implications in the care of stroke patients.

 

Based on a preliminary investigation conducted by the applicants (Dr. CP Lim and Dr. KS Tan), it is concluded that current resources pertaining to healthcare of stroke patients in Malaysia are limited.  In addition, there is a paucity of structured data and information concerning complications of stroke and accordingly, the proposed cross-disciplinary project is targeted to address this pressing issue.

 

In addition, Dr. KS Tan is a member of the Asia Stroke Advisory Panel (ASAP), and is involved in a project that endeavors to collect data and information pertaining to stroke patients in a number of Asian countries.  As a result, the outputs of this project can be readily extended to other ASAP participants in their countries.

 

As a generic approach will be adopted for developing the DSS, the resulting system can be adapted for diagnosis of other diseases such as heart attack which Dr. CP Lim has already embarked on in a research project in collaboration with Penang Hospital, Malaysia.

 

 

There has been a dramatic growth in the use of neural networks, an AI methodology, in medical applications [1].  The main thrust of work has been on the use of feedforward networks, primarily the Multi-Layer Perceptron and the Radial Basis Function network, as intelligent diagnostic and prognostic systems.  These networks are trained in a supervised fashion using a set of medical data pertaining to the problem at hand.  Normally, learning/training in these networks is an off-line process which consists of a training phase and a test phase using the collected data.  This approach is useful only when the data environment is stationary, and provided that training data are sufficiently representative.  This is because, during training, information (provided by the training data) about the problem is encoded by the adjustment of network weights.  After validating the network performance in a test phase, the network is put into operation and no further weight adaptation, or learning, takes place.  When the network is presented with a previously unseen sample, there is generally no built-in mechanism for the network to recognize the novelty.  In order to accommodate new information, the network will have to be re-trained using the new sample, together with all previous samples.  This drawback, suffered by most neural network models, arises from the so-called stability-plasticity dilemma [2].  The dilemma poses a series of questions: how can a learning system remain plastic or adaptive in response to significant events, and yet remain stable in response to irrelevant events?  How can a system adapt to new information without corrupting or forgetting previously learned information [2-3]?

 

To combat the stability-plasticity dilemma, ART-based neural networks have been proposed.  ART-based networks are incremental learning systems that self-organize and self-stabilize in response to an arbitrary sequence of patterns in non-stationary environments [2].  Each ART network includes a novelty detector that measures against a threshold the similarity between the prototype patterns stored in the network and the current input pattern.  When the similarity criterion is not satisfied, a new node can be created in the network with the input pattern coded as its prototype pattern.  Thus, the number of nodes grows with time, subject to the similarity criterion, in the process of learning a good network configuration incrementally.  As different tasks demand different capabilities from the network, this dynamic network architecture and autonomous learning methodology thus avoids the need to pre-specify a (static) network size or to re-train the network in non-stationary environments.

 

In this project, a specific type of supervised ART model, i.e. Fuzzy ARTMAP (FAM) [4], will be employed.  In addition to the ability of on-line, incremental learning, FAM offers an extra feature─a synthesis of neural networks and fuzzy logic.  This integration brings the low-level processing of neural networks and the high-level reasoning of fuzzy logic into a common framework to facilitate the development of intelligent systems.

 

The proposed FAM-based DSS is built on the successful advances made by the applicant (Dr. CP Lim) in which FAM has been further enhanced so that it is now able to operate in probabilistic environments, and to attain, on-line, the classification results predicted by Bayes’ decision theory [5-7].  Properties of FAM [8-10] and a number of successful applications using ART-based systems and other neural network models have been studied and demonstrated [11-16].  In terms of decision support, this means that the system can safely carry on learning in situ whilst providing useful support.  For example, in a diagnostic situation, evidence in the form of expert knowledge, measurements etc. would be input to a system.  If this input excites a recognized category (due to previous learning) in the system, then a predicted diagnosis will be returned.  Update of the system’s knowledge can only be initiated if and when the diagnosis is confirmed by medical experts.  If the current input is not recognized by the system, then the user is so informed.  This new piece of information will be incorporated into the system’s knowledge base. Again, adjustment of knowledge will only be initiated upon confirmation of the diagnosis.  In either case, as long as the diagnosis remains unconfirmed, no adjustment of knowledge takes place.  Using this methodology, any decision making or diagnostic procedure where evidence is to be associated with a definite outcome is a potential application domain for such an autonomous learning system.

 

 

The proposed research methodology and activities are as follows, with a flow chart and a Gantt chart in Figures 1 and 2.

 

Stage 1:            DSS Algorithm Development & Data Collection[1]

1.1                   Development of Fuzzy ARTMAP-based computational algorithms;

1.2                   Discussions with medical expertise, data collection and pre-processing

Outcome:         A robust computational model and code and a database of stroke patients

 

Stage 2:            MIR Control Structure Development

2.1                   Development of web-based control structure for flexible data manipulation (e.g. to facilitate incorporation of confirmatory or follow-up data);

Outcome:         A web-based control structure of MIR for data and information storage and analysis

 

Stage 3:            Software System Integration and Improvement

3.1                   Integration of outcomes from stages 1 and 2;

3.2                   Testing and debugging of the integrated system;

Outcome:         An integrated software system for stroke diagnosis

 

Stage 4:            Software System Evaluation and Implementation

4.1                   Assessment of the MIR integrated with the DSS using the stroke database;

4.2                   Derivation of necessary refinements

Outcome:         A validated web-based MIR with DSS for stroke diagnosis

 

 

It is envisaged that medical practitioners and researchers will be able to access anonymous medical records with heuristic diagnostic rules from the MIR via the Internet.  The information obtained will facilitate clinicians to apply the most effective curative and rehabilitative regimes to enhance quality of healthcare of patients, especially for those in poor and remote areas where infrastructure and medical expertise are scarce.  The absence of a web-based repository, such as MIR will be a deprivation for people in rural areas to enjoy quality medical and healthcare services that have been established in urban cities.

 

The MIR and DSS can be used as a resource, which contains up-to-date healthcare procedures and information, for continuing education and training of clinicians and medical practitioners.  In addition, the proposed project can be integrated into the tele-medicine flagship application under the Multimedia Super Corridor (MSC) project spearheaded by the Malaysian government.

 

Stage 5:            Project Consolidation

5.1                   Consolidation of all research findings and report writing

Outcome:         Detailed documentation and report for the proposed project