COMPUTER-ASSISTED DECISION SUPPORT SYSTEMS FOR MEDICAL APPLICATIONS

M. Frize, M. Stevenson
Dept. of Electrical Engineering, University of New Brunswick, PO Box 4400, Fredericton, NB, E3B 5A3

B. Nickerson
Faculty of Computer Science, University of New Brunswick P.O. Box 4400, Fredericton, NB, E3B 5A3

F.G. Solven
Dr. Everett Chalmers Hospital, P.O. Box 9000, Fredericton, NB, E3B 5N5

ABSTRACT

The application of case-base reasoning and neural network techniques to medical databases can be powerful techniques to aid physicians in making decisions about the management of their patients, in various types of medical units. Case-based reasoning allows to match cases (patients) using a predetermined hierarchical structure, a single patient parameter (text or numeric), to similar parameters contained in a clinical patient database. The output produces a group of patients which can be selected by matching exactly on certain characteristics or they may be matched "as closely as possible" using a gradient of patient properties. Neural network analysis is a pattern recognition technique which uses a training set of patient data (text or numeric) to seek mathematical relationships between various subsets of patient parameters. The discovered relationships learned from the training set are then applied to estimate desired outcomes. In this application to the ICU, the outcomes selected were mortality, length of stay, and hours of ventilation.

KEYWORDS: Medical informatics, case-based reasoner, artificial neural networks.

INTRODUCTION

The quick tempo of illness in critically ill patients has spawned numerous testing (monitoring) technologies which are evolving rapidly (Gardner, 1990; Zaloga, 1990) and generate large volumes of information. However, these monitors and hospital information systems are often unconnected, and generate many separate outputs. Recent literature discusses the usefulness and drawbacks of neural networks and expert systems as a means of collating this information to aid in clinical decision making (Chernow, 1990; Misiano, 1990). This need, coupled with the limitations of current severity of disease classification systems, such as the APACHE (Knaus et al., 1986), TISS, SEPTIC INDEX etc., attests to the importance of developing an integrated approach to process this amount of information. Most neural network and expert system developments in medicine, as decision- aid tools, have narrow applications.

The case based reasoning and neural network techniques developed by the Medical IDEAS Research Group are using various medical databases with a view to generalise the theory so that the techniques can have wide applications in the future. These techniques are currently being assessed with regards to their specificity, sensitivity and predictive values, involving a comparison of real clinical data, and data outputs from the measurement system developed.

METHODOLOGY

The research approach selected was to respect closely the manner in which physicians and nurses work in the patients units of interest. When a patient seeks medical attention, the physician interacts with the patient through the constructs of the medical model, an ancient ethical and intellectual code for the physician and provides structured interview and examination techniques directed towards establishing a diagnosis and a management plan. The diagnosis is usually stated as a pathophysiologic condition. Until a definitive diagnosis is established, often overlaying multiple chronic illnesses, the physician deals with a working diagnosis. Implicit in this approach are the concepts of a differential diagnosis, investigative plan and treatment plan. The differential diagnosis consists of a list of other likely pathologic entities which may explain the patient's condition. The investigative plan deals with the uncertainty in the differential diagnosis and invokes testing strategies to "rule out" or "rule in" the conditions suspected in the differential diagnosis. With the treatment plan, the patient's need for comfort is attended and an attempt is made, using surgery and/or medications, to return the pathologic state to the normal physiologic state. The modern interaction between the physician and the patient is often supplemented, especially in hospital practice, by testing systems (e.g. body fluid analysis, imaging, electrical signal analysis). Testing systems are used to reduce the level of uncertainty during the process of establishing the diagnosis and to monitor the effectiveness of the medical interventions (therapy).

A large medical database of over 2000 intensive care patients containing 98 fields of clinical and administrative information on patients admitted to the ICU at Dr. Everett Chalmers Hospital since 1988 was collected prospectively, with some retrospective chart review. In the current system, up to seven medical diagnoses and multiple procedures can be recorded, with auxiliary space for free-form comments. Significant events in the ICU course of the patient and complications that occur can also be noted. Current developments include a new screen for parameters concerning coronary artery and valve surgery patients. Another project consists in extending the case-based reasoning technique to a temporal study of rheumatoid arthritis patients (Gaskin & McKinnon, 1995).

The ICU database was integrated into the 'smart' system simulation and an early version of the case-based reasoner used ART-IM software (Automated Reasoning Tool for Information Management as the matching engine. The case-based reasoner used a modified Rete match algorithm to develop a match score based on string, word, character, or number matching. This algorithm is faster than conventional matching algorithms and allows (partial) matches or mismatches to be done on a string of characters, or on similar words in a string of words. Character matching uses trigrams to accommodate misspelled words. Number matching allows for the match weight to be scaled proportionately, based on a triangular function of the particular patient field. Each of these methods of matching and the match weights associated with them were selected in direct consultation with physicians to set the default values. The current system allows physicians to 'fine tune' any of the matching field weights to a patient's specific characteristics or to desired weight scaling (Frize et al.; Taylor et.al., 1993). This flexibility is important, since the same procedures performed on individual patients may have varying degrees of significance. As more information is known on the index patient case and entered into the program, the matched cases also change. When a patient has been discharged from the unit, the patient history is automatically added to the patient database. This allows the accuracy and quality of the data to improve with time and with use. The new prototype, IDEAS for ICUs (Intelligent Decision Aid System for Intensive Care Units) blends the rule-based ability along with the knowledge of medical experts on the team and issues warnings depending on the outcome of the matched cases.

Case-based reasoning is based on a weighted relationship of parameters developed in conjunction with physicians. It compares the ten closest matching cases to the current patient case. A three week clinical trial was done with the case-based reasoner in the Chalmers Hospital intensive care unit. More comprehensive trials are being planned for the systems in Winnipeg, Saskatoon and Fredericton in the 1996. In addition to using this technique with ICU patients, a new research endeavour is now applying this technique (with appropriate variations) to a longitudinal database of rheumatoid arthritis patients. The objective is to develop a technique for assessment of various patient outcomes.

Another aspect of the project was to assess the applicability and usefulness of various pattern classification techniques to process critical care information. During the initial development, the feed-forward back propagation artificial neural network (ANN) was chosen because of its ease of implementation and past success on various classification tasks (Buskard et al., 1994; Frize et al., 1995). The performance of a neural network is affected by its architecture (the number of layers, the number of hidden units, etc.). Experimentation with various architectures led to the selection of a back propagation feed-forward ANN. The weights of the network were adjusted in such a way as to minimize the mean-squared error at the output of the network. Given a sufficiently large training set, a sufficiently large neural net should implement the optimal Bayesian classifier and error rate. In practise, the neural network error rate is usually higher than the Bayesian error rate due to the fact that the training set size and network size are limited and that the training algorithm may find a local minimum, instead of a global minimum.

This technique was used to estimate length of stay in the ICU and Intermediate Care areas, the duration of artificial ventilation requirements and estimation of mortality. A number of the input variables consist of the physiologic components of the APACHE scoring system, in addition to other variables such as sex and the unit from which the patient was transferred (admission source). The output predicting the various outcomes was compared to true patient data from the large database collected. Approximately two-thirds of the patient files from the large ICU database were used as a training set and the remainder as a test set to evaluate the performance of the classifier. A new ANN is currently under development using a coronary-artery bypass and valve surgery patients database. Outcomes, in this case, are similar to the ICU study, but other parameters will be added in the future (McGowan, 1996).

DISCUSSION AND RESULTS

The case-based and neural network systems collate a large amount of clinical information and presents it in simplified formats. Clinicians may thus use the system to find the group of the closest matching cases to their current patient case. This approach is similar to that used by physicians when they are thinking: "I have seen a patient like this" and provides instant recollection of past cases that may be relevant to the present case. In an intensive care unit (ICU) setting, for example, aspects of the history of the selected group may then be displayed graphically in terms of outcomes, such as mortality, length of stay, hours of artificial ventilation, procedures utilized and complications encountered. A graphical user-interface has been developed using Visual Basic in the MS-Windows environment. The interface is user-friendly and the medical staff on the research team have participated in every step of the development of its configuration.

Preliminary work with the neural network has already achieved an improvement in the accuracy of predicting mortality and length of stay. It has also produced the first results in estimating the duration of artificial ventilation (Frize et al., 1995). Clinical trials will assess the effectiveness and usefulness of both systems and whether they have an impact on length of stay, mortality and hours of ventilation, in a population similar to that found in the databases.

CONCLUSION

This research is meant to enhance the conventional medical model used by physicians for patient management in an Intensive Care environment. It uses previous patient experience recorded in a large clinical database with new patients and generates outputs which may be helpful for physicians in their decision-making. This approach may be applicable to other patient care environments and current efforts are focussed in generalising the tools and theory for wider applications in the future.

REFERENCES

Buskard, T., M. Stevenson, M. Frize and F. Solven (1994), Estimation of ventilation, length of stay, and mortality using artificial neural networks, Proc. of 1994 Canadian Conference on Electrical and Computer Engineering, (IEEE) Sept., 1994, Halifax, pp. 726-729.

Chernow, B. (1990), The bedside laboratory: A critical step forward in ICU care, CHEST, May supplement, 97 (5), 183S-184S.

Frize, M., K. B. Taylor, B.G. Nickerson, F.G. Solven, H. Borkar, V. Dunfield, A Knowledge- Based System to Assist Patient Management in an Intensive Care Unit, Proc. of the IMIA- IFMBE Working Conf., Aalborg, 1993, pp.156-159.

Frize, M., F.G. Solven, M. Stevenson, B. Nickerson, T. Buskard, and K. Taylor, (1995), Computer-assisted Decision Support Systems for Patient Management in an Intensive Care Unit, Medinfo'95, July, Vancouver, 1009-1012.

Gardner, R.M., M. M. Shabot, Computerized ICU Data Management: Pitfalls and Promises, Int. J. of Clinical Monitoring and Computing, vol. 7, 1990, pp.99-105.

Inference Corporation, Case-Based Reasoning in ART-IM, Version 2.5, Inference Corporation, 550 N Continental Blvd., el Segundo, California, 1991.

Knaus, W.A., E.A. Draper , D.P. Wagner, J.E. Zimmerman, An Evaluation of Outcome From Intensive Care in Major Medical Centres, Annuals of Internal Medicine, vol. 104, 1986, pp.410-418.

McGowan, H.C.E., M.H. Stevenson, M. Frize, The Need for Standardized Reporting of Medical Applications of Artificial Neural Networks, CMBEC, Charlottetown, June 26-29, 1996.

Misiano, D.R., M.E. Meyerhoff, M.E. Collison, (1990), Current and Future Directions in the Technology Relating to Bedside Testing of Critically Ill Patients, CHEST, 204S-214S.

Taylor, K.B., B.G. Nickerson, M. Frize, F.G. Solven, V. Dunfield, Use of Case-Based Reasoning to Assist Patient Management in an Intensive Care Unit, Proc. of the joint conference of COMP and the CMBES, Ottawa, 1993, pp.248-249.

Zaloga, G.P. (1990), Evaluation of Bedside Testing Options For The Critical Care Unit, CHEST, May supplement, 97(5): 185S-190S.