68

Review

J. Dent. 1992; 20: 68-73

Expert systems in dentistry. Past performance-future prospects

S. E. Stheeman, P. F. van der Stelt and P. A. Mileman Department of Oral Radiology, Academic Centre for Dentistry Amsterdam (ACTA), Amsterdam, The Netherlands

ABSTRACT Expert systems are knowledge-based computer programs designed to provide assistance in diagnosis and treatment planning. They assist the practitioner in decision making. A search of the literature on expert system design for medical and dental applications was carried out. It showed an increase in the number of articles on this subject. Between 1984 and 1991,608 articles have been published in medical journals and two in dental journals. Because it is likely that this development will influence dental practice in the future a critical review of medical literature on the topic has also been carried out. A number of general principles are described to give the dental practitioner some insight into how expert systems work A set of criteria have been formulated from the medical literature which expert systems should meet. These requirements are also applicable to dentistry and may be used to judge dental expert systems. In the last part of the paper the features of several dental expert systems developed in the past decade are described in the light of these criteria. It is concluded that in the future more attention should be paid to the development and evaluation of expert systems in the clinical setting. Only well-designed and properly evaluated expert systems can be expected to earn a place in everyday practice.

KEY WORDS: Expert systems, Computers, Review

J. Dent. 1992; 20: 68-73 (Received 18 April 1991; reviewed 28 June 1991; accepted 25 September 1991)

Correspondence should be addressed to: Dr S. E. Stheeman, Department of Oral Radiology, Academic Centre for Dentistry Amsterdam, Louwesweg 1, 1066 EA Amsterdam, The Netherlands.

INTRODUCTION

Medical expert systems which attempt to improve and support diagnostic decision making or treatment planning have been the subject of increasing attention in the last decade. Expert systems are knowledge-based computer programs designed to provide assistance in diagnosis and treatment planning. They assist the practitioner in decision making. In medicine, working expert systems have been in existence since the 1970s (Bramer, 1981). The development of these systems has become widely recognized as a research priority (Adams et al., 1986). Some problems concerning implementation of these systems however remain, such as their ease of use and the lack of commonly accepted agreement in clinical termin- ology used by the medical profession (de Dombal, 1987). At this time only a small number of medical expert systems are in use in clinical practice (Lucas and Janssens, 1985).

Modern medicine has become so complex that it is difficult for the doctor to keep up with all current developments (de Dombal, 1988a). There are great

@ 1992 Buttenvorth-Heinemann Ltd. 0300-57 12/92/020068-06

differences in diagnostic problem-solving techniques among medical practitioners (Taylor et al., 1971). The number and type of additional tests requested vary greatly among medical students near the end of their training (de Vries RobbC, 1978). Expert systems can retain and process more items of information and knowledge than human beings (Kinney, 1987). Large amounts of data can be integrated and processed by them correctly (Erdman, 1987). They are therefore an excellent way of making available to professionals, with a core of basic knowledge, large amounts of specialized knowledge about a particular field. In addition to acting as a reference, using an expert system can be an excellent way to manipulate and store patient information needed for diagnosis, treatment planning and evaluation.

Computer consultative aids may decrease the cost of health care because they are less expensive than human consultants and because they may also reach clinical decisions more quickly (Taylor et al., 1971) with less expensive supplementary diagnostic tests. They can also improve diagnostic accuracy because of their systematic

200

r

1984 1985 1986 1987

Year

1988 1989 1990

Fig, 1. Number of publications on medical and dental expert systems 1984-90. ??, Evaluation; 0, general.

and thorough approach (Erdman, 1987). Expert systems are capable of improving the quality of the clinical decision-making process (de Dombal et al., 1972; Gorry and Silverman, 1978).

From a review of the literature it is apparent that there is an increase in the number of publications concerning expert systems (Fig 1).

The use of expert systems in dentistry is likely to play an increasingly important role in the future. The working field of the practitioner is an always expanding one in which it is difficult to keep up with the constant flow of information. New techniques are constantly being introduced and procedures that are complex or were not possible in the past are now undertaken by the dentist in everyday practice. The field of the dental specialist is constantly shifting and referrals to the specialist can be expected to be an area of future innovation. The available knowledge about disease, diagnosis, treatment and sub- sequent prognosis is growing by the day, making the working field more specialized. More than ever there is a need for detailed information and fast information processing. Dentists are already adopting computerized administrative systems and the proportion of dentists presently using them is growing (Stheeman, 1989). It seems likely that dental expert systems in the future will combine administrative packages in such a way that they are available on demand in the surgery. However, at present, the lack of agreement on medical terminology makes the transfer of information between two or more programs very difficult. It is suggested that it is likely that the development and use of dental expert systems will increase substantially. It will be important for the dental practitioner to grasp how such systems work. The rest of this article will therefore review a number of articles about

Stheeman et al.: Expert systems in dentistry 69

dental expert systems that were published in the past 5 years. First however a few general principles will be explained concerning the use of existing expert systems in medicine.

EXPERT SYSTEMS IN MEDICINE

The knowledge needed to develop an expert system is often ill defined (Shortliffe, 1986). To collect the detailed information needed to build an expert system is often a tedious and time-consuming task (Shortliffe, 1986). In addition, knowledge in medicine is of an imprecise nature and involves a degree of uncertainty, for example, in the relationship between symptoms and diagnosis. In order to design expert systems that can make decisions using this imprecise knowledge many approaches have been developed. These fall into two major groups. First there are methods that calculate what diagnosis is statistically the most likely or what decision is best in the given circumstances. Secondly, symbolic reasoning can be used to reach a conclusion. In practice a combination of both approaches is often used.

Probabilistic algorithms

Epidemiological information about many diseases has been published. This information is numerical and can easily be used in conjunction with a computer.

One of the best known ways of dealing with probabilities is known as analytical decision making and it is based on ‘Bayes’ theorem’ (Bayes, 1763). Using this method the probability that a given patient is suffering from a particular disease can be calculated from the probability that symptoms present in the patient coincide with that disease given that the prevalence of disease and symptoms in the population are known. For each disease built into the program this probability can be calculated. The assumption is made that the patient suffers from the disease with the most likely outcome. This is a commonly used approach in decision-making analysis in dentistry (Mileman et al., 1986; Tulloch and Antczak-Bouckoms, 1987; Maryniuk et al., 1988).

The main difficulty in applying ‘Bayes’ theorem’ is the large amount of data needed to determine conditional probabilities used in the formula (O’Neil, 1990). The prevalence for each disease and the probabilities of their occurrence in a patient with certain symptoms must be known. The method faulters when one or more diseases are characterized by only a few features (de Vries and de Vries Robbe, 1985). Therefore the Bayesian approach works best with well-defined problems. Expert systems using decision analysis have been shown to be difficult to transfer from one patient population to another (de Dombal et al., 1981). Poor user acceptance may occur in spite of a highly valid system (de Dombal, 1988b). Examples of such programs are described by Nugent et al. (1964) Wameretal. (1961)andFitzgeralderal. (1966).De

70 J. Dent. 1992; 20: No. 2

Dombal(1988a) however published the experiences of 13 years of successful use of an expert system for the computer-assisted diagnosis of acute abdominal pain based on the method described.

The ‘pattern match’ approach is another probabilistic method sometimes used in expert system design. Here the set of symptoms and clinical findings from a patient is compared with the set of symptoms and findings typical for a disease. This is done for every disease known to the program. The disease or condition that best matches the patient’s symptoms and clinical findings is considered to be the most likely diagnosis. Compared to the analytical decision-making approach, this method is easier to implement because it only requires knowledge concerning the symptoms not their prevelance. It is used in a few programs developed in dentistry either in a role supporting probabilistic calculations (White, 1989) or as the main part of the expert system (Abbey, 1987).

Symbolic reasoning

It is possible to represent expert knowledge as a set of rules. These have the basic form: IF situation THEN action/conclusion. Rules of this kind are called production rules.

Because of the imprecise nature of medical knowledge these rules are often combined with certainty factors. Such a rule might have the following form: IFsymptoms THEN possible/probable/most likely cause or another possible form of the same information is: IF symptoms THEN cause 0.7. In the latter example of a production rule the certainty factor is 0.7 (see Fig 2). It represents a degree of belief in the conclusion about the cause of the symptoms. A combination of symbolic reasoning and quasi- probabilistic reasoning has been developed in which a hypothesis may be proven, rejected or accepted (Shortliffe and Buchanan, 1975).

The main disadvantage of this kind of reasoning is that these rules often are not explicit in the medical literature. It often takes experts to define them (de Vries and de Vries Robbt, 1985) which is very time consuming. It is also very difficult to make the rules ‘watertight’. A number of high- performance systems have been developed using these types of rules. Perhaps the best known of them all is MYCIN which advises physicians about selection of antimicrobial therapy (Shortliffe, 1974).

In addition to defining knowledge in the form of production rules there are also systems designed in which the factual knowledge in the computer program is made completely independently of the general reasoning strategies. The philosophy behind such a system is that the knowledge can be made available in this way for many different tasks. An important example of a system with such a broad scope is INTERNIST-I (Pople et al., 1975; Miller and Pople, 1982).

Not only the medical data but sometimes also the data entered by the user of the expert system is of an imprecise nature. For instance, suppose when an expert system asks

IF: 1) The infection is primary-bacteraemia, and 2) The site of the culture is one of the sterile sites,

and 3) The suspected portal of entry of the organism is

the gastrointestinal tract, THEN: There is suggestive evidence (0.7) that the identity

of the organism is bacteroides.

Fig. 2. An example of a production rule from MYCIN.

the user whether the heart rate of a patient is ‘elevated’, each user may have a different rate which he calls ‘elevated’. Therefore the answer will be imprecise or ‘fuzzy’. This kind of imprecise data can be quantified by using so called ‘truth values’ as percentages for any possible answer from the user. This might mean, for example, that a heart rate of 110 beats per minute is considered definitely elevated (meaning that the truth value will be 100 per cent). At a heart rate of 85 however, the truth of it being ‘elevated’ might be 50 per cent because it is in a range of what may be considered to be ‘normal’ heart rate. When the answers of the user have been quantified they can be used in decision making, for example, diagnosis. This kind of reasoning is called ‘fuzzy logic’ (Zadeh, 1983) and it is used in one of the dental expert systems which will be discussed below.

What demands should be made of expert systems?

From the experiences with expert systems in medicine it is possible to define a set of requirements for the use of expert systems in everyday dental or oral surgery practice, as follows:

1. The system should be easy to use and respond quickly (Timpka, 1987).

2. The system should easily tit into the current clinical diagnostic and therapeutic strategy without demanding a change of approach from the user in diagnostic or therapeutic problem solving.

3. The system should be robust enough to be able to function with incomplete information; it should be able to render an ‘intermediate result’ if not all the data is available, and suggest what missing information should be obtained and in what way.

4. The user should always be in control of the process of consulting an expert system (de Vries RobbC et al., 1989).

5. The system should have a clear explanatory facility (Erdman, 1987; Teach and Shortliffe, 1981; Wallis and Shortliffe, 1982) preferably explaining its conclusions in terms of knowledge instead of showing the rules that led to its conclusions (de Vries Robbe et al., 1989); the program could suggest literature references about the subject under consideration.

6. The specificity and sensitivity of the results of applying the program to a diagnostic problem should be as high as possible.

7. The knowledge used should be generally accepted among experts in the field.

Stheeman et al.: Expert systems in dentistry 71

8. The effectiveness of the system should have been evaluated by potential users from different geographical settings.

9. The system should be designed to be able to be integrated into administrative systems in current use (Timpka, 1987) or be an easy to implement reference system.

EXPERT SYSTEMS IN DENTISTRY

In dentistry several papers have been published concern- ing expert systems at various stages of development (Table I). Most of the common approaches to the development of those systems. described earlier, have also been used to develop knowledge-based programs in dentistry. Several expert systems published in the dental literature will be discussed below.

An expert system for oral diagnosis

Abbey (1987) published an introductory article on the ‘periapical radiolucency’ module of this medical diag- nostic expert system. The module is part of a system written using software developed by Weed and Hertzbeg (1983). According to Abbey a clinician can develop a tailored expert system concerning any diagnostic problem using the specifically manufactured software. The data- base contains four lists of possible causes, findings, comment and literature references, and management options.

In an expert system built with the software the diagnostic process is reportedly controlled by several sets of questions. The user can answer them in any order. The author describes as an example a part of the system called ‘periapical radiofucency’. This part is able to produce a differential diagnosis for a periapical radiolucency. The causes are ranked in descending order by the number of observed findings, so the best matching cause is searched for, given the findings.

An important feature of this expert system design is that a list of comments and literature references to the dental literature are added to it. The user can therefore receive information about the reliability of a finding or of important aspects of the diagnosis. In his article, Abbey does not report on the evaluation of the program in a clinical setting.

An expert system for oral radiographic diagnosis

ORAD (an acronym for Oral Radiographic Diagnosis) (White, 1989) is an experimental computer program designed to aid oral radiographic differential diagnosis. The program uses analytical decision making to calculate the most likely diagnosis regarding the patient’s symp- toms. In order to do so the program contains info~ation describing the prevalence of 97 diseases and conditions.

In addition to using decision analysis to calculate a diffe~ntial diagnosis of the disease present for the patient, the program also calculates a direct ‘pattern match’ of symptoms and signs. When both the outcomes of the probability and ‘pattern match’caIculations are high for a certain condition, then it is considered additionally likely that the patient may indeed be suffering from it. The program does not offer any explanation of its conclusions.

White (1989) describes four patient cases on which the program performed well. There is no report on a general evaluation of the program.

An expert system for orthodontic advice

This computer program has been written to guide the general dental practitioner with limited orthodontic knowledge in orthodontic problem solving. Its purpose, according to the authors (Sims-Williams et al., 1987), is to provide the dentist with specialist knowledge to judge whether or not he or she should seek the opinion of a consultant in orthodontics. In addition some information about treatment options is provided by the system.

The authors used fuzzy logic in the development of their expert system. When using the program the clinician answers a set of questions with rather inexact possible answers. Therefore the program uses truth values to work with the uncertainty involved in the answers. The expert system reportedly designs a treatment plan and defines the reasons for choosing it in clear language, which would make it easy to work with.

The authors conclude that the program’s advice closely matches the advice that would have been given by the expert upon whose knowledge it was based. They do not provide an evaluation of their system or report whether the knowledge that was used is generally accepted by other experts in the field.

An expert system for endodontic diagnosis

In 1983 Wyman and Diehl reported on the development of an expert system for endodontic diagnosis. It uses 19 symptoms and test results. Four diagnoses are possible, namely: reversible or irreversible pulpitis, necrosis or a healthy pulp.

The program was tested on 38 patients who had been referred for endodontic examination. The program recommended endodontic treatment in 83 per cent of the cases in which treatment was actually performed. The program’s level of agreement was 100 per cent for the diagnosis of necrosis or irreversible pulpitis and 85 per cent for the diagnosis of reversible pulpitis. There was no difference noticed in the level of agreement when molars or non-molars were considered.

The authors of the article clearly state that the program only considers pulpal disease, so that it really assumes only few solutions to a diagnostic problem.

72 J. Dent. 1992; 20: No. 2

Table 1. The described expert systems in dental diagnostics and treatment planning

Application Methodology Evaluation

Oral diagnosis’ Oral radiographic

diagnosis2

Orthodontic treatment advice3

Diagnosis of pulpal disease4

Treatment planning of dental trauma5

Modified pattern match Not reported Bayesian statistics Four cases are reviewed; no

evaluation of the outcomes of the program in comparison with clinical data is reported

Fuzzy logic Not repotted

Bayesian statistics 38 patients for agreement of diagnosis and treatment planning clinical vs computer

Not reported Not repotted

‘Abbey (1987). 2White (1989). Sims-Williams et a/. (1987). 4Hyman and Diehl (1983). 6Hyman and Doblecki (1983).

An expert system for the diagnosis of dental trauma

To assist emergency personnel in the diagnosis and the planning of initial treatment of patients with dental trauma, Hyman and Doblecki (1983b) developed an expert system. They hoped to achieve considerable time savings in the initial treatment by its use. This could have meant an improvement in the initial care for dental trauma by treatment providers who are not familiar with dental injuries.

The expert system considered 20 different injuries to teeth and their supporting structures. The program suggests treatment that could be provided by a dentist having a fully equiped dental surgery. The authors suggest that this could be altered to reflect limitations in the level of training and equipment of the user.

No information about the approach used in the expert system was given in the article, nor was there a report of an evaluation of the system.

FUTURE PROSPECTS

The research that has been done in the development of expert systems in dentistry looks promising. At this moment there are more diagnostic dental expert systems than systems that assist in treatment planning. It is likely that there will be a place in dental practice for diagnostic expert systems as well as expert systems that support treatment planning.

The approaches that can be choosen for building decision-supporting systems are various. Depending on the task the expert system is designed to perform, one method may however be more appropriate than another. When looking at a field like oral radiographic diagnosis for instance, the complexity of facets of the knowledge

used in the diagnostic process and the present lack of sufficient hard epidemiological data suggest that symbolic reasoning together with the use of certainty factors is the approach of choice. On the other hand, the decision analysis approach is more suitably applied to better defined fields, for instance dental trauma. Careful selection of the approach required for each specific type of problem will pay off in the long run.

Important progress could be made by the development of expert systems that have an elaborate explanatory function. Explanations improve the acceptance of the expert system’s conclusions by the user and might be an important tool in postgraduate education.

Expert systems should be designed to fit into existing diagnostic set-ups, preferably with knowledge that is commonly accepted among international experts in the field. Ninety per cent of all medical expert systems have not been independently evaluated for performance in clinical settings (Lundsgaarde, 1987). To be able to rely on the results of any given expert system, it should have been properly evaluated. We therefore conclude that evaluation in a clinical setting is a precondition for the implementation of expert systems in dental practice. Only well-designed and properly evaluated expert systems can be expected to earn a place in everyday practice.

Acknowledgement

We wish to thank Dr P. E. Zanstra from the Research Group Medical Information and Decision Science, University Hospital Groningen, The Netherlands for his thorough review of the introductory part of this paper.

Stheeman et a/.: Expert systems in dentistry 73

References Abbey L. M. (1987) An expert system for oral diagnosis. J.

Dent. Educ. 51,475-480. Adams I. D., Cham M., Clifford P. C. et al. (1986) Computer-

aided diagnosis of abdominal pain: a multi-centre study. Br. Med. .I. 293, 800-804.

Bayes T. (1763) Philos. Trans. R. Sot. Land. [Biol.] 53, 370. Bramer M. A. (1981) A survey and critical analysis of expert

system research. In: Hammersley P. (ed), Information Technology for the Eighties. British Computer Society, pp. 488-515.

de Dombal F. T. (1987) Computer-aided decision support- the obstacles to progress (editorial). Methods Znf: Med. 26, 183-184.

de Dombal F. T. (1988a) Computer-aided diagnosis of acute abdominal pain: the British experience. In: Dowie J., Elstein A. (eds), Professional Judgement. London, Cambridge University Press.

de Dombal F. T. (1988b) Computer support for medical decision making. Rev. Med. Chir. Sot. Med. Nat. Zasi 92, 137-144.

de Dombal F. T., Leaper D. J., Staniland J. R. et al. (1972) Computer-aided diagnosis of acute abdominal pain. Br. Med. J. 2, 9-13.

de Dombal F. T., Staniland J. R., Clamp S. E. et al. (1981) Geographical variation in disease presentation. Does it constitute a problem and can information science help? Med. Decis. Making 1, 159-169.

de Vries P. H. and de Vries Robbe: P. F. (1985) An overview of medical expert systems. Methods Znf: Med. 24, 57-64.

de Vries Robbe (1978) Medische besluitvorming-een aanzet tot formele geneeskunde. PhD Dissertation, University of Groningen.

de Vries RobbC P. F., Zanstra P. E. and Beckers W. P. A. (1989) Verschillende redeneermechanismen in MEDES. In: Siklossy L. and Bakker R. R. (eds), Proceedings NAIC ‘89. Academic Service Schoonhoven, The Netherlands, pp. 59-67.

Erdman H. P. (1987) A computer consultation program for primary care physicians. Impact of decision making model and explanation capability. Med. Care 25, ~138-147.

Fitzgerald L. T., Overall J. E. and Williams C. M. (1966) A computer program for the diagnosis of thyroid function. Am. J. Roentgenol. 97, 901-905.

Gerry G. A. and Silverman H. (1978) A computer program that considers clinical responses to digitalis. Am. J. Med. 64,452-460.

Hyman J. J. and Diehl M. C. (1983) A dental trauma diagnostic program. In: Proceedings of the 7th Annual Symposium on Computer Applications in Medical Care. New York, IEEE.

Hyman J. J. and Doblecki W. (1983) Computerized endodontic diagnosis. J. Am. Dent. Assoc. 107, 755-758.

Kinney E. L. (1987) Medical expert systems. Who needs them? (editorial). Chest 91, 3-4.

Lucas P. J. F. and Janssens A. R. (1985) Medische expertsystemen: hulpmiddel bij diagnose en therapie. Ned. Tijdschr. Geneeskd. 129, 160-164.

Lundsgaarde H. P. (1987) Evaluating medical expert systems. Sot. Sci. Med. 24, 805-819.

Maryniuk G. A., Schweiter S. 0. and Braun R. I. (1988) Replacement of amalgams with crowns: a cost-effectiveness analysis. Community Dent. Oral Epidemiol. 16, 263-267.

Mileman P. A, Vissers T. and Purdell-Lewis D. J. (1986) The application of decision making analysis to the diagnosis of approximal caries. Community Dent. Health 3, 65-81.

Miller R. A. and Pople H. E. (1982) INTERNIST-I, an experimental computer based diagnostic consultant for general internal medicine. New. EngJ. J. Med. 307, 468-476.

Nugent C. A., Warner H. R., Dunn J. T. et al. (1964) Probability theory in the diagnosis of Cushing’s syndrome. J. Clin. Endocrinol. 24, 621-627.

O’Neil M. (1990) An Introduction to Medical Expert Systems. The Tutorial from the 6th Annual Meeting on Expert Systems in Medicine, London 1990.

Pople H. E., Myers J. D. and Miller R. A. (1975) DIALOG: a model of diagnostic logic for internal medicine. ZJCA17.5, 848-855.

Shortliffe E. H. (1974) MYCIN: A Rule-based Computer Program for Advising Physicians Regarding Antimicrobial Therapy Selection. PhD dissertation, Stanford, Stanford University.

Shortliffe E. H. (1986) Medical expert systems-knowledge tools for physicians. West. J. Med. 145, 830-839.

Shortliffe E. H. and Buchanan B. G. (1975) A model of inexact reasoning in medicine. Math. Biosci. 23, 351-379.

Sims-Williams J. H., Brown I. D., Matthewman A. et al. (1987) A computer-controlled expert system for orthodontic advice. Br. Dent. J. 163, 161-166.

Stheeman S. E. (1989) Geautomatiseerde praktijkvoering. Een inleiding. Ned. Tijdschr. Tandheelk. 96, 200-202.

Taylor T. R., Aitchinson J. and McGirr E. M. (1971) Doctors as decision-makers: a computer-assisted study of diagnosis as a cognitive skill. Br. Med. J. 3. 35-40.

Teach R.i. and Shortliffe E. H. (1481) An analysis of physical attitudes regarding computer-based clinical consultation systems. Comput. Biomed. Res. 14, 542-558.

Timpka T. (1987) Knowledge-based decision support for general practitioners: an integrated design. Comput. Methods Programs Biomed. 25,49-60.

Tulloch J. F. C. and Antczak-Bouckoms A A. (1987) Decision analysis in the evaluation of clinical strategies for the management of mandibular third molars. J. Dent Educ. 51, 652-660.

Wallis J. W. and Shortliffe E. H. (1982) Explanatory power for medical expert systems: studies in the representation of causal relationships for clinical consultation. Methods Znf: Med. 21, 127-136.

Warner H. R., Toronto A. F., Vesay L. G. et al. (1961) A mathematical approach to medical diagnosis: application to congenital heart disease. JAMA 177, 177-183.

Weed L. L. and Herzberg R. Y. (1983) The use and construction of problem-knowledge couplers, the knowledge editor, knowledge network and the problem-oriented medical record for the microcomputer. In: Proceedings of the 7th Annual Symposium on Computer Applications in Medical Care ‘83. New York, IEEE Pub Svc.

White S. C. (1989) Computer-aided differential diagnosis of oral radiographic lesions. Dentomaxiflofac. RadioJ. 18, 53-59.

Zadeh L. A. (1983) The role of fuzzy logic in the management of uncertainty in expert systems. In: Fuzzy Sets and Systems. Amsterdam, North Holland (Elsevier Science Publishers), pp. 199-227.