COMPUTER-ASSISTED DECISION SUPPORT SYSTEMS FOR PATIENTMANAGEMENT IN AN INTENSIVE CARE UNIT

M. Frize, M. Stevenson, T. Buskard and K. Taylor
Dept. of Electrical Engineering, 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, E3B5N5

B. Nickerson
Faculty of Computer Science, University of New Brunswick, P.O. Box4400, Fredericton, NB, E3B 5A3

ABSTRACT

The application of the intelligent monitoring techniques of case-basedreasoning and neural network analysis to physician decision making concerningpatient care in an Intensive Care Unit (ICU) is described. Case-based reasoningoffers a model for quickly matching-using a predetermined hierarchicalstructure-a single patient's parameters (text or numeric) to similar parameterscontained in a clinical database. The output produces a group of patientswhich may be set to match exactly on certain characteristics and may alsobe set to match "as closely as possible" on a gradient of patientproperties. Clinicians may thus use the system to find the group of theclosest matching cases to their current patient. Aspects of the ICU historyof the selected group may then be displayed graphically (e.g mortality,length of stay, hours of ventilation, procedures utilized and complicationsencountered). Neural network analysis is a pattern recognition techniquewhich uses a training set of patient data (text or numeric) to seek mathematicalrelationships between various subsets of patient parameters. The discoveredrelationships from the training set are then applied to estimate the outcomes(e.g. mortality, length of stay, hours of ventilation) of new patients.The effects of these intelligent monitoring techniques are scheduled tobe tested in a field trial held in a regional referral centre ICU.

INTRODUCTION

When a patient seeks medical attention, the physician interacts withthe patient through the constructs of the medical model. The medical modelis an ancient ethical and intellectual code for the physician and providesstructured interview and examination techniques which are directed towardsestablishing a diagnosis and a management plan. The diagnosis is usuallystated as a pathophysiologic condition (e.g. heart failure) due to an anatomicabnormality (e.g. a myocardial infarction). Until a definitive diagnosisis established about the presenting problem (which, in adult medicine oftenoverlays multiple chronic illnesses), the physician deals with a workingdiagnosis. Implicit in the working diagnosis are the concepts of a differentialdiagnosis, investigative plan and treatment plan. The differential diagnosisis a list of the other likely pathologic entities which may explain thepatient's condition. The investigative plan deals with the uncertaintyin the differential diagnosis by invoking testing strategies to "ruleout" or "rule in" the conditions listed in the differentialdiagnosis. With the treatment plan, the patient's need for comfort is attendedand an attempt is made, using surgery or medications, to return the pathologicstate to the physiologic.

The modern interaction between the physician and the patient is oftensupplemented, especially in hospital practice, by testing systems (e.g.body fluid analysis, imaging, electrical signal analysis). Testing systemsdeal with uncertainty and are used most often as part of the differentialdiagnosis (to increase or decrease the likelihood of a disease process)and to monitor treatment. If the patient is healthy or the diagnosis andprognosis are known, then no testing need be done. Even if a definitivediagnosis (e.g. myocardial infarction) is established, the natural history(i.e., how the illness unfolds over time) is often uncertain for any onepatient. Thus, there is a need for testing systems to deal with the uncertaintyin the diagnosis and management of patients.

Our medical research group has focused on the application of the intelligentmonitoring techniques of cased- based reasoning and neural network analysisas an aid to patient management in the critical care environment. We havepositioned these techniques to work with the traditional medical modelas seen in Figures 1 and 2.

The quick tempo of illness in critically ill patients has spawned numeroustesting (monitoring) technologies which are evolving rapidly [1,2] andgenerate large volumes of information. However, these monitors and hospitalinformation systems are often unconnected, uncoordinated and generate confusingoutputs. Recent literature discusses the usefulness and drawbacks of neuralnetworks and expert systems [3-5] as a means of collating this informationto aid clinical decision making [6,7]. This need, and the limitations ofcurrent severity of disease classification systems, such as the APACHE[8], TISS, SEPTIC INDEX etc., attests to the importance of developing anintegrated approach to process this type of information. Previous developmentsin neural networks and expert systems [3-5] as decision aid tools in healthcare, are normally designed for narrow medical applications or for a particularpathology [1, 8-12].

The objective of this research project was to investigate the developmentand use of a knowledge-based system and of a neural network to providean enhanced "intelligent bedside monitoring" capability for usein a critical care area. Using conventional statistical techniques involvinga comparison of real clinical data and data outputs drom the measurementsystem developed, the patient information processing and testing techniquesdeveloped are currently being assessed with regards to their specificity,sensitivity and predictive values.

METHODOLOGY

A large medical database of over 2000 intensive care patients was madeavailable to our group [13]. The database contains 98 fields of clinicaland administrative information on patients admitted to the ICU at DECHsince 1988. Data collection was primarily prospective, with some retrospectivechart review. Since its inception, null entries have been recorded as distinctfrom zero entries, to ensure data integrity. Up to seven medical diagnosesand multiple procedural information can be entered, with auxiliary spacefor free form comments. Significant events in the ICU course of the patientand complications that occur are also noted.

This database was integrated into the 'smart' system simulation. A veryearly investigation into the development of a case-based reasoner leadto the creation of a prototype CaIC (Case-based reasoner for IntensiveCare) [13], using ART-IM software (Automated Reasoning Tool for InformationManagement). The case-based reasoner uses a modified Rete match algorithmto develop a match score based on string, word, character, or number matching.The Rete match algorithm is faster than conventional matching algorithms.String matching either matches or mismatches on a string of characters.Word matching matches on similar words in a string of words. Charactermatching uses trigrams to accommodate misspelled words. Number matchingallows for the match weight to be scaled proportionately, based on a triangularfunction of the particular patient field. Each of these methods of matchingand the match weights associated with them were selected in direct consultationwith the physician on the research team, to arrive at the preset defaultvalues. We have explored the concept of different preset weights dependingon the general patient category upon admission (e.g. postop, trauma, cardiac,respiratory, endocrine, or other). It is possible for the physician usingthe system to 'fine tune' any of the matching field weights to a patient'sspecific characteristics. This flexibility was felt to be important, sincethe same procedures performed on individual patients may have varying degreesof significance. As more information is known on the index patient caseand entered into the program, on the index patient case, the matched casesalso change. When a patient has been discharged from the ICU, the patienthistory is automatically added to the patient case base. This allows theaccuracy and quality of the data to improve with time and use.

Based on this work, a more improved prototype case-based reasoner, IDEASfor ICUs (Intelligent Decision Aid System for Intensive Care Units) [14,15],was developed. The knowledge-based system uses ART-IM (Automated ReasoningTool for Information Management) software [16]. ART-IM has both case-basereasoning and rule-base reasoning capabilities. It is the blending of therule-based ability along with the knowledge of medical experts on the teamthat was the foundation of the development of a decision aid tool whichinteracts with and enhances the medical model. IDEAS for ICUs issues warningsdepending on the outcome of the matched cases. For example, a strong warningwill be issued if a match is made on a patient who died in the ICU andhas a high match score with the index patient.

Another aspect of the project was to assess the applicability and usefulnessof various pattern classification techniques to process critical care information.During the initial development of this project, the feed-forward backpropagationneural network was chosen because of its ease of implementation and pastsuccess on various classification tasks [16,17,18]. This technique wasused to estimate length of stay in the ICU and Intermediate Care areas,duration of ventilation requirements, and mortality. The inputs to thesystem were selected to be the physiologic components of the APACHE scoringsystem together with sex of the patient and admission source from withinthe hospital. The engineering system output which predicts patient statusand outcome was compared to true patient data from the large database collected.

The performance of a neural network is affected by its architecture(the number of layers, the number of hidden units, etc.). Thus, some experimentationwas required in order to determine the appropriate architecture. The backpropagationtraining algorithm adjusts the weights of the network so as to minimizethe mean-squared error at the output of the network. It has been shownthat minimization of the mean-squared error results in the optimal Bayesianclassifier [18]. This is the objective of the backpropagation trainingalgorithm. Thus, given a sufficiently large training set, a sufficientlylarge neural net should be capable of implementing the optimal Bayesianclassifier and of achieving the Bayesian error rate. In practise, the neuralnetwork error rate will be higher than the Bayesian error rate due to thefact that the training set size and network size are limited and that thetraining algorithm may find a local minimum instead of a global minimum.Approximately two-thirds of the patient files were used as a training setand the remainder as a test set to evaluate the performance of the classifier.

DISCUSSION AND RESULTS

The case-based and neural network systems collate large amounts of clinicalinformation and present it in simplified formats. The current output ofthe case-based system provides outcome assessment and is based on a weightedrelationship of parameters developed in conjunction with physicians. Agraphical user interface is being developed using Visual Basic in the MS-Windowsenvironment. The interface is user-friendly and the medical staff on theresearch team have participated in every step of the development of itsconfiguration. Preliminary work with the neural network has already achievedan improvement of 10 percent in the accuracy of predicting mortality (from80 percent with APACHE II to 90 percent with our current system, givenour hospital population) and the preliminary results in estimating lengthof stay and duration of ventilation show a 70 percent classification rate[20].

The strengths and weaknesses of each of the two engineering approachesare currently under study. The project intends to use each in an optimalmanner and link them to provide a more complete approach to the processingof medical information in the ICU.

The clinical trials results will be available by the end of the year.We will prospectively examine the effect of the system upon length of stay,mortality and hours of ventilation in a population of patients in a regionalreferral centre ICU.

CONCLUSION

This research is an enhancement of the conventional medical model usedby physicians for patient management in an Intensive Care environment.It compares previous patient experience held in a large clinical databasewith the new patient and generates an output which maybe used to aid physiciandecision making. This system positions itself between the medical modeland usual testing systems and represents a new class of "intelligentmonitoring instrumentation" for critical care information analysis.In the future, this approach maybe applicable to other patient care environments,such as neonatal care, cardiac units, neuro-intensive care and the operatingroom.

REFERENCES

1. Gardner R.M., Shabot M.M., Computerized ICU data management: pitfallsand promises. Int. J. of Clinical Monitoring and Computing, vol. 7,1990, pp.99-105.

2. Zaloga, G.P. (1990) Evaluation of bedside testing options forthe critical care unit. CHEST, May supplement, 97(5): 185S-190S.

3 Alonso-Betanzos A., et al., A Connectionist Approach to PredictAntenatal Outcome. Proc. of the 14th Ann. Int. Conf. of the IEEE EMBS,Paris, 1992, pp.1004-1005.

4. Fu LiMin, A Hybrid Medical Expert System. Proc. of the 13thAnn. Int. Conf. of the IEEE EMBS, Orlando, 1991, pp.1290-1291.

5. Hudson D.L., Cohen M.E., Banda P.W., A Hybrid Expert System forDevelopment of Diagnosis and Treatment Plans. Proc. of the 13th Ann.Int. Conf. of the IEEE EMBS, Orlando, 1991, pp.1284-1285.

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

7. Misiano, D.R., Meyerhoff, M.E., Collison, M.E. (1990) Currentand future directions in the technology relating to bedside testing ofcritically ill patients. CHEST, 204S-214S.

8. Knaus W.A., Draper E.A., Wagner D.P., Zimmerman J.E., An evaluationof outcome from intensive care in major medical centres. Annuals ofInternal Medicine, vol. 104, 1986, pp.410-418.

9. Arcay B., Moret V., Balsa R., Hernandez C., Physical and functionalintegration system for intelligent processing and priorization of variablesin an ICU. Proc. of the Ann. Int. Conf. of the IEEE EMBS, Seattle,1989, pp.1993-1994.

10. Barro S., Ruiz R., Mira J., More than classical patient monitoringsteps towards intelligent approaches. Proc. of the Ann. Int. Conf.of the IEEE EMBS, Seattle, 1989, pp.1995-1996.

11. Teres D., Brown R.B., Lemeshow S., Predicting mortality of intensivecare unit patients. The importance of coma. Critical Care Medicine,10(2), 1982, pp.86-95.

12. Thibault G.E., Mulley A.G., Barnet G.O., Goldstein R.L., Reder V.A.,Sherman E.L., Skinner, E.R., Medical intensive care Indications, interventionsand outcomes. The New England Journal of Medicine, 302(17), 1980, 938-942.

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

14. Frize M., Taylor K.B., Nickerson B.G., Solven F.G., Borkar H., AKnowledge-Based System for the Intensive Care Unit. Proc. of the 15thAnn. Int. Conf. of the IEEE EMBS, San Diego, October 28-31, 1993.

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

16. Schalkoff, R. (1992) Pattern Recognition: Statistical, structuraland neural approaches. John Wiley & Sons.

17. Inference Corporation, Case-Based Reasoning in ART-IM. Version2.5, Inference Corporation, 550 N Continental Blvd., el Segundo, California,1991.

18. Wan, E. (1990) Neural network classification: A Bayesian interpretation.IEEE Transactions on Neural Networks, 1(4): 303-305.

19. Wong D.T., Knaus W.A., Predicting outcome in critical care:thecurrent status of the APACHE prognostic scoring system. Can. J. Anaesth.,38(3), 1991, pp.374-83.

20. Buskard T., Frize M., Stevenson M., Solven. F. Development ofan Artificial Neural Network for the ICU. The 6th Annual Meeting ofthe Medical Research Network of New Brunswick. 1994.