Developing Microprocessor Based Expert Models for Instrument Interpretation

DEVELOPING MICROPROCESSOR BASED EXPERT MODELS

FOR INSTRUMENT INTERPRETATION

Sholom M. Weiss, Casimir A. Kulikowski

Department of Computer Science Rutgers University New Brunswick, New Jersey

ABSTRACT

We describe a scheme for developing expert interpretive systems and transferr ing them to a microprocessor environment. The scheme has been successfully implemented and tested by producing a program for interpreting results from a widely used medical laboratory instrument a scanning densitometer. Specialists in the field of serum protein electrophoresis analysis provided the. knowledge needed to build an interpretive model using EXPERT. a general production rule system for developing consultation models By constraining a few of the structures used in the general model. it was possible to develop procedures for automatically translating the model to a specialized application program and then to a microprocessor assembly language program Thus. the model deveiQpment can take place on a large machine. using established techniques for capturing and conveniently updating expert knowledge structures. while the final interpretive program can be targeted to a microprocessor depending on the application. Our experience in carrying out the complete process illustrates many of the requirements involved in taking an expert system from its early development phase to actual implementation and use in a real world application.

I INTRODUCTION

Most knowledge-based medical consultation systems developed during the 1970's were relatively large-scale experimental prototypes [ 1. 2. 3l Their advice on diagnostic and treatment problems typically involved approximate reasoning over a space of many interrelated hypotheses. characteristically supported by hundreds of observations linked to them by various types of reasoning rules By adopting symbolic reasoning methods w ith more powerful representations than the traditional mathematical decision-making schemes. these knowledge-based systems produced results that were generally easier to analyze. explain and update than those from more conventional systems. Human engineering features were often stressed as an important me.ans of enhancing the interaction with the expert systems. Successful clinical experience with many of these systems has been reported in pilot demonstration projects. yet none is in routine clinical use at present. Both technical and social factors contribute to the difficulties of introducing expert systems into the every day practice of med1cine One often cited technical factor IS the slow rate of manual data entry required by most of the larger systems. This problem is minim1zed for applications where most of the data can be read d1rectly off a clinical instrument. and only few items need be entered manually. The commercial availability and use of automated electrocardiogram interpretation programs (using traditiOnal Robert S. Galen Department of Pathology Columbia University New York. New York algorithmic techniques) supports this point. Regardless of the methods used in constructing a knowledge base. or its complexity. instrument-derived interpretations are more likely to be accepted because they can be seen as extensions of the instrument And since many advanced medical instruments are already microprocessor controlled. it may be possible to add an interpretive module that enhances the performance of such a device at relatively little extra cost In this paper we briefly describe how we were able to accelerate the development of interpretive software for a widely used laboratory instrument. the scanning densitometer. We did this by automatically producing a computer translation of an expert model for serum protein electrophoresis interpretation. developed on a mainframe computer. into a microprocessor assembly language vers1on. The translation methods have been generalized so that this process can be repeated for EXPERT [4] models in any domain. with a few restrictions on the types of knowledge structures used.

By taking this approach. we have demonstrated that knowledge engineering methods from expert systems can be used to full advantage in producing an effective model. which can then be transferred with ease to a microcomputer.

II METHODS

Several general purpose schemes for building consultation systems have evolved from work on the earlier more specific domain-dependent systems. Two such schemes that were originally designed for representing medical consultation problems in particular are the EXPERT and EMYCIN [5] systems. Both provide built-in control mechanisms operating over specif ic types of production­ rule models. The consultation program of EXPERT is primarily event-driven. while that of EMYCIN is predominantly goal directed.

The EXPERT system has been used in building a number of expert medical consultation models (mainly in ophthalmology, rheumatology. and endocrinology) and pilot prototypes in several non- medical areas (spectroscopy interpretation. car repair . hazardous spill management. and oil well-log interpretation).

The process of model design and transfer that we used in developing the microprocessor based expert model for serum protein electrophoresis interpretation involved the following tasks·

• Specification of the knowledge base using
• EXPERT;

• Empirical testing with several hundred cases;
• Refinement of the knowledge base by the expert.
• Further cycle of testing with additional cases.
• and review by independent experts;
• Test of the final model on the large machine;
• Automatic translation of the EXPERT model to a specialized program and a microprocessor assembly language program;
• Interfacing of assembly language model with instrument.

This last step requires detailed knowledge of the instrument In this application the manufacturer interfaced the interpretive program to the existing program for printing instrument readings.

Figure 1 illustrates the types of conclusions reached by the interpretive system. and the type of rule used in reasoning. The most significant restriction on the type of production rules used in the model was to limit the use of confidence measures to 3 values representing confirmation. denial. or unknown status. In applications of this type. it should be noted that the strategy of questioning is not a signif icant task because most of the information will be obtained directly from the instrument. In building the EXPERT model we simulated this situation by entering the values of certain key features of the instrument signal (Figure 2). that are currently given as a digital output by the instrument. These features include peaks of the waveform and areas under certain segments of the waveform. A few items (patient identification. age, and $Ome waveform features that are more easily scanned by the technician) are entered manually.

The serum protein electrophoresis model required several stages of refinement over a period of six months. with the aid of one principal expert and comments and suggestions from the independent experts. We began with a relatively small and simple model. having 1 0 conclusions and a production rule for each. After the first cycle of revision we had about 25 conclusions and 50 rules which included many for handling interactions among the hypotheses. The current model has 38 conclusions and 82 production rules. Its performance on 256 test cases covering the full spectrum of conclusions is 1 00% acceptable to our experts. They expect differences of oponion on the amount of detail included in the present set of conclusions. but feel that covering infrequently found problems would detract from a model that is to be dosseminated widely. An option for allowing users to add a written record of their own opinions on such unusual cases has been provided in the final microprocessor implementation

Ill CONCLUSIONS

The completed microprocessor version of the interpretative serum protein electrophoresis model may not look much different than if it had been hand coded directly on the assembly language of the microprocessor. or translated from an algorithmic language There is.

nevertheless. a fundamental difference. With our system.

we can rapidly produce new versions of the microprocessor program from our high level EXPERT model in response to any changes suggested by the experts or resulting from future empirical analysis and clinical tests in the field. In contrast. considerable effort would usually be required to recede directly on a microprocessor. Besides. the original expert derived model is also very different than one produced by more traditional methods. Our conclusions and intermediate hypotheses were developed in such a way that they include not only diagnostic considerations but also prognostic. treatment. and future test selection decisions for motivating their use The large amount of experimentation that took place with the model as it went through its cycles of testing and modification could only be carried out on a larger system. with adequate facilities for analyzing many cases. and knowledge engineering tools for changing the model. A recently published version of an interpretive model in this domain developed with very traditional programming techniques. shows a contrasting sparsity in diagnostic statements [6] . In addition. the conclusions of this model appear to be overly specific given the nature of the supporting data. Thus. while programs of this type may be initially simple to implement. they do not incorporate the elements of expert reasoning that are essential to a clinically helpful program.

In conclusion. the work reported here is a novel illustration of the requirements needed in taking an expert system from an early developmental phase to actual implementation and use in the real world. S!JCh applications can lead to the increasing acceptance of expert systems in medicine and other domains where similar problems can be

found.

ACKNOWLEDGMENTS

This work was supported in part by grant RR643 of the NIH DRR-BRP. Technical assistance was provided by

Helena Laboratories. Beaumont. Texas.