Welcome to ehealthpedia.org, Our Health Informatics Wiki

From ehealthpedia - Our Health Informatics Wiki

Computer –assisted Medical Decision –making

INTRODUCTION

Decision –making by the clinician in the management of his patients is a highly intellectual activity which involves, (i) his skill in gathering and evaluating new information about the patient, (ii) his ability to readily recapitulate the information he has already logged in the patient’s record, and (iii) his ability to effectively utilize the large body of medical knowledge which expresses the relationship between the data describing each individual patient and the diagnostic, prognostic and therapeutic options available for managing the patient’s problem optimally. The computer can facilitate and improve the clinician’s performance of each of these three tasks. The patient-clinician dialogue (adopted from Komaroff) is depicted in Fig. 10.1. The large amount of existing medical knowledge and the rapid growth of that knowledge during the last quarter of this century have resulted in a situation wherein most physicians find it increasingly difficult to assimilate all the information which would be useful in making optimal clinical judgments. Specialization and super-specialization provide a partial solution to this problem, but the rapid evolution of technology and clinical research makes it difficult even for the specialist to keep up. In the light of this ‘information explosion’, it is not surprising that empirical studies have demonstrated that physicians do not always make optimal decisions. This problem extends across all aspects of medical decision-making from diagnosis to patient management.

For the last 25 years, the idea has been advanced that the Computer –assisted Medical Decision making (CMD) system might provide a solution to much of the problem created by this information explosion. The idea of computer programs that can directly assist the doctor with decision –making is at once intriguing and disconcerting; intriguing because of the potential to improve medical care and medical education, not so much through the discovery of new bio-medical knowledge as through a more effective application of that which already exists; disconcerting because of the potential for abuse and alteration to the practice of contemporary medicine. In an article in the British Medical Journal (BMJ) in 1999, Lawrence Weed emphasized the need to use properly designed computer software tools to enable the practicing clinician to perform a proper combinational analysis for appropriate decision –making—gathering all relevant data about a patient and linking it with a medical knowledge base. He produced problem-knowledge couplers to enhance the performance of the individual care-giver and to protect against error. Information tools are superior to the human mind as a device for retrieving and processing knowledge.

A CMD system can be defined as an interactive computer system that directly assists doctors or other healthcare professionals with clinical decision-making tasks. The system is intended to support (not replace) doctors, complementing their natural abilities to make judgements with the computer’s vast memory, reliability and processing capabilities. Other motivations for developing CMD systems include their potential educational value, their use as an ‘intelligent’ interface to medical databases, and their role as models of the diagnostic reasoning process. Interaction of the clinician with the computer programmer has focused attention on the clinicians’ thinking process itself. When the clinician says that ‘common things occur commonly’, the computer programmer uses Bayes’ theorem of conditional probabilities. When the clinician uses a branching logic according to the symptom present or absent, the computer programmer uses Boolean logic, set theory and symbolic logic. The logic circuits of digital computers are organized according to the operational principles of Boolean algebra, and operations are performed upon sets, like cluster analysis and multivariate analysis. The human brain is unsurpassed in its ability to perceive, focus, think, analyze, imagine and create, but it is greatly limited in its ability to store a large collection of facts permanently, to recall the facts instantaneously and precisely, and to handle multiple variables at a time. It is in these areas that the computer will supplement the human brain and vastly improve the clinician’s performance in decision –making.

GENERAL MODEL OF CMD

Dr. James Reggia has described a general model of CMD system as depicted in Fig. 10.2. The input is typically a description of some specific patient (age, sex, symptoms and signs, past medical history, etc.) and the output is useful information about than patient (e.g., appropriate screening tests, diagnosis, therapy plan, etc.). The CMD system itself contains two basic components: (i) a knowledge base, and (ii) an inference mechanism. The knowledge base is a collection of encoded knowledge that is needed to solve problems in some particular medical area. The inference mechanism is a program that, given a case description, uses the information in the knowledge base to generate new information about the case. While these two components are viewed here as being conceptually distinct, they have been interwoven in some CMD system.

Doctors may interact with CMD systems in various roles. Individuals who use a complete CMD system to assist them with problem solving are referred to as users. Expert physicians who provide the information for a CMD system’s knowledge base are called medical experts or knowledge base authors. Finally computer scientists who design and develop the supporting software for CMD systems are referred to as knowledge engineers. In the past, these roles usually corresponded to separate individuals. Knowledge engineers typically serve as intermediaries between medical experts and the CMD system under construction, directly helping the experts to express their problem –solving knowledge in a form suitable for machine processing. This has been a difficult and time consuming task. The fundamental problems in the creation of CMD system are, (i) knowledge representation –how to do represent human knowledge in terms of data structures that can be processed by machine? (ii) inference generation –how do we use these abstract data structures to generate useful information in the context of a specific case, and (iii) knowledge acquisition –how do we translate human knowledge as it currently exists in medical textbooks, journal articles, clinical databases, and the minds of physicians, into the abstract representations that are being used? These and other problems have proven to be much more difficult to handle than early CMD enthusiasts had anticipated.

VARIOUS APPROACHES TO DECISION -MAKING

Murphy has characterized the clinician’s approach to diagnosis as being of four types; (1) the exhaustive approach ; (2) the Gestalt approach; (3) the algorithmic or multiple branching approach; and (4) the hypothetico –deductive approach.

The exhaustive approach is taught to the medical student but rarely practised by the expert clinician except in difficult problem cases. Expert clinicians also use this approach when dealing with problems outside their domain.

The Gestalt approach is that of a unified configuration and interpretation of the data elements that cannot be derived from the sum of its parts. The semantic pattern recognition approach is closely related to that of the Gestalt. Humans are superb at pattern recognition, but when attempting to emulate human expert behaviour, it is difficult for them to identify and codify the features of patterns that are so critical to recognition. Gestalt includes the notions of intuitions and hunches which are difficult to identify and quantify.

The algorithmic approach provides a sequential and categorical or deterministic guide to decision and action. Here it must be noted that though the decisions made in the algorithmic approach are deterministic, the design of the algorithm may be based on probabilistic features. Algorithms provide a useful skeleton to illustrate the major features of decisions and may be used in pattern recognition schemes and vice versa, and also to both develop hypotheses and to test them. Algorithms for hypothesis development are among those that have been classified in other areas as pattern recognition algorithms and profile interpretation algorithms.

Both the Gestalt and the algorithmic approaches appear to be involved in the rapid and aggressive hypothesis generation characteristic of the clinician when confronted with familiar problems. The Gestalt approach operates here when a diagnostic interpretation appear to be unambiguous and almost instantaneous. This shades into an algorithmic approach when a sort of pre-compiled list of routine questions may also be required to effect important or common discriminations between alternative diagnostic hypotheses.

The hypothetico-deductive approach is the one most used by the expert clinician based on his mental model of the patho-physiological events through are produced. Thus, if a symptom is produced via a number of patho-physiological mechanisms between which it is important to differentiate, the clinician may well ask about two or three symptoms or obtain laboratory data likely to be associated with a particular mechanism. His choice of a particular mechanism is probably based on his impression of its infrequency in the sub-population with which he is dealing (‘common things occur commonly’) and also of the seasonal characteristics of the disease. Thus, phenomenological features (as distinct from the patho-physiological) are also invoked. His early choice of associated questions may also depend on the seriousness of the association of a finding with a particular mechanism. It is the elegant explanation of unusual patterns via extensive understanding of basic mechanisms that sets apart the virtuoso performance of an expert clinician.

One of the great potentials of computer –aided decision –making is that the routine pattern recognition and management systems that have been devised by expert clinicians in different fields will be available to the primary physician, forming a bridge between the routine and the more difficult cases. It will raise the primary physician’s performance to the level of the expert.

It is clear that the clinician uses various approaches which are not mutually exclusive. He may dynamically switch between approaches and can promptly switch out of his customary and routine mode into a more structured and probabilistic approach. The designers of computer expert systems face the task of more precisely defining the circumstances in which each approach is to be used. In fact, currently developing systems tend to use more than one approach in each system as appropriate to each phase and the sub-goal of a problem –solving task. Burke suggests the intimate interplay of categorical and probabilistic approaches in his flow charts of the diagnostic and management processes as depicted in Figs. 10.3 and 10.4.

COMPUTER –ASSISTED DECISION SUPPORT SYSTEMS

For ease of discussion, it can be said that the CMD systems adhere to the following approaches:

• Categorical Approaches: Deterministic, algorithmic, flow-chart or protocol –based. They provide clear –cut guides to action based on clear-cut criteria. Indeed, the concept of a protocol as a clinical algorithm arose out of the experience of research workers with computers and information science. • Probabilistic Approaches:

Statistical decision approaches and statistical inference; Bayesian approaches; Discriminant analysis; Multivariate analysis; Clinical analysis –case-based reasoning; and Exploratory analysis, etc.

All the above are derived from the same perspective –to develop predictive power through the analysis of past data to support future decisions.

• Artificial Intelligence Approaches Production rule systems based on first order predicate calculus –conditional rules, IF/THEN;

Cognitive models based on generalized set covering (GSC) theory;

Frame descriptions; Semantic networks; Hypothesis and test (abduction); and Artificial Neural Networks (ANNs).

ALGORITHMIC METHODS

An algorithm is a step-by-step set of instructions on how to accomplish a task. The algorithmic approach is successful when the patho-physiology is clearly understood and categorical decisions can be made on the basis of reliable laboratory information. Examples of the successful clinical use of this approach are given in Table 10.1. The use of clinical algorithms or protocols offers a way of both studying the cost –effectiveness of health practices and implementing a system which encourages conformance to practices that have been deemed medically sound and cost-effective.

1. If vaginal discharge, itch, irritation, dysuria occur, do pelvic exam, Pap smear and GC culture. 2. If dysuria and/or urinary frequency, record presence of incontinence ! Incontinence; ! Vomiting or nausea; ! Fever in past two days; ! History of bladder or kidney stones; and ! Consult physician before treatment. 3. If urine symptoms, do urinalysis and urine culture. 4. If analysis shows 2 + bacteria, 20 WBC or greater, treat with sulfa unless allergy is present. 5. If abnormal vaginal discharge, do wet prep. 6. If Monilia prep. Positive, treat with mycostatin. 7. If trichomonas prep. Positive, treat with metroindazol. 8. If abnormal discharge but neither 6,7 positive, treat with sulfa unless allergy is present.

 Table 10.1. Protocol and Decision Rules for UTI/Vaginitis in Women

Elements of a Protocol

A protocol begins with a defined medical problem. For the particular problem, the protocol does the following six things:

1. It defines which patients with that particular problem can be managed appropriately by the user of the protocol. 2. It specifically indicates those history questions, elements of physical examination and laboratory tests which need to be collected in order to manage the problem. 3. It includes branching logic rules that individualize the clinical data that is collected according to particular characteristic of the patient, such as age, sex, past medical history, current medication and characteristics of the patient’s complaint. Therefore, the same clinical information is not collected on every patient. 4. The protocol designed for non-medical users indicate very specifically those clinical findings which are serious enough to require referral to or consultation with a doctor. 5. The protocol includes precise rules for arriving at a diagnostic impression and for making management decisions such as prescribing drugs. 6. The protocol can be filed in the record and serve as the progress note for the encounter.

Display of protocol logic can be in the form of flow charts; these are found in increasing numbers in medical journal articles designed for clinical readers. Such systems have also been used in training non-medical personnel to care for ambulatory patients with a variety of symptoms and conditions. Protocol logic may also be displayed as a host of decision rules. Each rule corresponds to one node or branching box in a flow chart. Thus a large number of rules of needed to reproduce all the logic in a good-sized flow chart. An example of decision rules for UT/vaginitis in women is given in Table. 10.1.

When the total number of combinations of clinical findings is not very large, and when the management process is not highly sequential, that is, when the process divides conveniently into a data-collection stage and an action –stage rather than several contingent action stages, then a decision table can be used.

Paper-based protocols have the advantage of being inexpensive and easy to disseminate to a large number of dispersed users. The increasing availability of low cost time-sharing computer services in a variety of setting has made it possible to present the sequential logic of the protocol to user. Since no protocol can be completely comprehensive, it is necessary to build in specific referral or consultation points which alert the user that the case in question is one that the protocol is unable to deal with and that the user should seek consultation from an appropriate clinician. A protocol whose scope is broader than the abilities of the user is potentially dangerous, pushing the user into unknown areas beyond the point at which referral should have necessarily taken place. The scope of the protocol and precise specification of the complaints or conditions that the protocols is designed for, and for which patients it is appropriate (age, sex, usual health status), must be clearly known to the user.

Protocol for Training Healthcare Providers

Protocol have been extensively used to train non-medical healthcare workers in a limited set of specific tasks to be mastered by the trainee. Protocol as training tools also have the virtue of emphasizing the procedural or operational aspects of medical care. A properly designed protocol system includes documentation describing and with a given protocol is to be used (that is, on which patients and with what problem). Once a protocol is chosen, that instrument then guides the user in deciding what questions should be asked of the patient, what examination performed, and so on. This is in contrast to conventional training in which the student is given body of information about disease and then left to his own devices regarding the best procedure or strategy to be followed in deciding which disease is present.

Protocol for Quality Assurance

Since a protocol represents a logically complete set of the pre-set criteria for dealing with a clinical problem, it is natural that protocols have found use as the basis of a number of quality assurance systems in healthcare. Computers facilitates scrutiny of formatted medical records to examine whether the protocol rules were followed. As a result of such an exercise, either the protocol or the clinician’s behaviour, or both, are modified to reduce the discrepancy between the care actually given and that recommended by the protocol. Such an exercise also helps in cost containment, for example, decreased ordering of a specific test, or increased ordering of a lower cost generic drug.

The representation of a sequential strategy as a flow chart often provides a graphic display of that strategy which is clearer and easier to comprehend than a textbook description would be. Hence one finds increasing use of flow charts in medical journal articles. These provide the clinician with an immediate consultation for a large number of situations in which the clinicians may be uncertain as to how to proceed.

How to Write a Good Protocol

Many Practitioners seek a protocol for the diagnosis and management of acute illness in primary ambulatory care. The following steps are recommended by Pass, Komaroff and Ervin in writing a protocol:

1. Identify the patient for whose care the protocol is intended (for example, an adult male presenting to a primary care facility) 2. Identify a cluster of related complaints to be handled by a single protocol. 3. Identify the disease or conditions which may cause those complaints. 4. Characterize each disease or condition in terms of

(i) Its likelihood of occurrence in a primary care setting (ii) Its treat ability, if found (iii) The Seriousness of its consequences (iv) The clinical skills and resources need to identify it

5. On the basis of the above considerations, decide which of the possible causes of the patient’s problem the protocol

(i) Cannot afford to miss (ii) Should refer to an expert or specialist (iii) Is treatable by the protocol user, or (iv) Can be ignored

6. Identify the symptoms, signs and laboratory tests most likely to be useful in diagnosis the conditions which must be identified, if present 7. Write the rules specifying the circumstances under which each of these data is to be collected, each possible diagnosis is to be made, each possible therapy given ad when a referral from the protocol is to be made 8. Translate these rules into a sequential series of steps to be taken by the user, including all appropriate branching logic 9. Test, validate and correct the protocol, through (i) Initial peer review (ii) Preliminary trials (iii) Prospective randomized controlled trials, comparing the protocol and its user with the conventional mode of care 10. Document the protocol, explaining the medical rationale for each step of the algorithm and describing the clinical skills required to use it 11. Disseminate the protocol and supporting documents to the users.

If the protocol is for use especially in acute illness, by non-medical personnel, formal validation trials are essential to show the safety and efficacy of the protocol, essentially on medico-legal grounds. The Ambulatory Care Project (ACP) formed in 1969 at the Beth Israel Hospital in Boston shows wide-ranging experience with computer –driven interactive protocols for several chronic diseases to guide health assistants (high school graduates) with a brief four-week training. An evaluation of these efforts was reported by Komaroff in 1974. Physicians accepted the system as reflected by their attendance rate at the clinic. Increase through ness in collecting clinical data in the protocol system led to an increase in the recognition of new pathology. Time –motion studies showed a 20 per cent saving of the clinicians’ time with this protocol system.

Evaluation of the acute complaint protocol system has shown that briefly trained health assistants, guided by a protocol could independently evaluate and treat 68 per cent of the men coming to a US Army Clinic with acute genito- urinary infection systems. Headache, abdominal pain and respiratory infections have also been tackled satisfactorily and achieved significant time-saving for physicians without sacrificing the quality of care.

Essex observed in Tanzania that flow chart protocols for common problems seen in rural clinics performed well. These was 98 per cent agreement among those using the chart to examine independently the same patient and there was disagreement among only 6 percent of the students and the experienced physicians. Protocols designed to manage specific common problems of specific regions or entire countries, such as those designed by Essex and Hirschhom, seem to have great potential for developing countries. Briefly trained health workers using protocols could dramatically increase the quality to care delivered in the Third World. Patient Care magazine has adopted the logical flow chart format to display recommended clinical strategies to its physician readers. Most other major medical journals have also published articles which employ this method to present suggested strategies. JAMA has published a series of articles on the optimal use of clinical laboratories in which protocols for a recommended sequences of test ordering are presented in the logical flow chart format.

Probabilistic Approaches to Decision-Making In 1968, Dr. Henry Wagner proposed a model of the entire diagnostic process based on the sequential application of Bayes’ Theorem of conditional probabilities. He emphasized that medical diagnosis was probabilistic in nature and was hence eminently suited for applying the probability theory:

“Every question that the physician asks while obtaining a medical history, every manoeuvre that he performs in the physical examination, and every subsequent laboratory procedure that he orders, should be selected because of the likelihood that the new fact will alter the estimate of probability that the patient has a particular disease or diseases. With the availability of microcomputer, ayes’ Theorem seems to be an idea whose time has come.”

Bayes’ Theorem

The essence of Bayes’ Theorem of inverse probability is:

Di = P (Sj/Di) xP (Di) Sj  ? [ P(Sj/D) xP (D)]

This equation states that the probability P that a patient with a given syndrome Sj has a particular disease Di is directly proportional to the probability of the occurrence of his syndrome in that disease, P(Si/Di) multiplied by the prevalence of that disease P (Di), and inversely proportional to the probability of the occurrence of his syndrome in all other disease D multiplied by the prevalence of those disease P(S/D) x PD.

Since it is usually not possible nor practical to analyse a large enough population of patients to derive estimates of P(S/Di) for every possible patient descriptor S, the assumption is often made in practice that the individual facts SK that together form S are independent (the known occurrence or absence of one sign or symptom does not change the probability that some other sign or symptoms will be found to be present ). Using the assumption, it follows that

[P(S/Di) = P(Sj/Di) x P (S2/Di) x …..P (Sm/Di)

In this assumption is used, the knowledges base does not require the potential astronomically large set of 2m probabilities P(S/Di) for each Di (we presume the Si to be binary), but only the m probabilities P (SK/Di) of each individual symptom SK for each disease Di. The Bayesian system has been used for diagnosis and assessment of prognosis. Figure 10.5 shows a logical system of medical diagnosis based on this approach.

F. de Dombal and associates developed a Bayesian CMD system for the diagnosis of acute abdominal pain that was much more accurate than the physician ( 90 percent correct diagnosis by computer vs 80 percent correct diagnosis by the physician). Had the computer’s prediction about the occurrence of acute appendicitis been used than that of the attending clinician, not only would far fewer unnecessary operations have been performed, but also fewer cases of actual appendicitis would have been put under observation and delayed from receiving appropriate surgery. Certain characteristic of human thinking may be detrimental to management of the complementary problems of uncertainty and complexity. The human tendency is to over-estimate low probabilities and to under-estimate high probabilities. Hence deDombal considers computers in some area of medicine better at discriminating among the common causes of disease than doctors.

In 1975, Rifkin and Hood evaluated a Bayesian approach for the interpretation of ECG stress-testing. They examined the accuracy of predicting the angiographic evidence of coronary artery disease from a quantitative analysis of the degree of exercise-induced ST segment depression and found the predictive value assessed by a Bayesian approach to be quite accurate.

Diamond and Forester have developed a program, which is now commercially available, based on an extensive review of published data on clinical and laboratory manifestations of coronary artery disease. Diamond comments: “We may be standing at the threshold of a perceptual upheaval in medicine. Our perception of apparently simple, categorical questions of diagnostic judgment can be expanded from a lean dimensionless point into a rich three –dimensional whole.”

Homer Warner’s pioneering effort in the diagnosis of congenital heart disease was based on a Bayesian approach. For every patient, 60 different indicator (history, physical findings, laboratory data, etc.) have to be provided. Obtaining all this information in every case is often clinically unreaslistic. In 1983, Zaggoria and Reggia used the Bayesian approach for predicting prognosis in stroke, as an example, let us assume that a study of 1,000 patients with stroke found three factors were ago of the patient, type of stroke and severity of the stroke. 16 percent of the patient, type of stroke and severity of the stroke. 16 percent of the patients were classified as having a good recovery, 59 per cent as fair and 25 per cent as having poor recovery. Using this knowledge base, one can see the possible outcomes and their probabilities. Thus, for example of the patients who eventually had a good outcome, 25 per cent were aged 60 years of more and 75 per cent were below 60. Ten per cent had an intra-cerebral haemorrhage, 20 percent had a thrombotic infarction, 40 per cent had an embolic infarction and 30 per cent had a sub-arachnoid haemorrhage; 78 percent had a mild stroke, while 22 per cent had a moderate to severe stroke. Using this specific knowledge base, and given a new 58-years-old patient with a mild embolic stroke, a Bayesian CMD system would estimate the probability of good outcome as 31 percent, a fair outcome as 68 per cent, and a poor outcome as 1 percent, by substituting appropriate probabilities in the Bayesian formulae mentioned above.

An important question is whether data regarding prior probabilities used in Bayesian CMD generated from clinical experience at one center may be useful at another location, with a different geographic, ethnic and socio-economic milieu. Some evidence about the transportability is provided by Dombal (1977). Further studies of this nature are obviously needed from different countries, including India and China.

The Bayesian approach makes the assumption that the patient has one disease at a time, which is not valid in many real-life situations. In 1980, Ben Bassat, et.al., proposed a modification of the Bayesian approach to handle multiple disorders. A Bayesian approach typically assumes that a patient’s signs and symptoms are independent. Such an assumption can result in loss of performance comprising the theoretical claim of minimal mis-classification made for this method. A variety of solutions have been proposed to handle this problem (D. Fryback, 1978). The sequential application of Bayes’ Theorem can be used (Gorry and Barnett, 1968).

Sequential Bayes

A sequential Bayes calculation is needed when more than one finding is used to determine if a patient has a disease. Bayes’ formula is applied to the first test becomes the prior probability for the second test. For example, in a patient presenting with chest pain, we are considering the probability of acute myocardial infarction. If the disease prevalence is 3.7 per cent and infarct chest pain in positive in 75 per cent of these with the disease and given a false positive in two percent of those without the disease, then:

P (D) = 0.037 P(ND) = 0.963 P(T/D) = 0.75 P (T/ND) = 0.02

Bayes’ formula updates the probability of the disease in a patient with the presence of the symptoms as follows:

(0.75) (0.037) P(D/T) = = 0.59 or 59 per cent (0.75) (0.037) + (0.02) (0.963)

The interpretation of this calculation is that only 3.7 in every 100 hospital patients have acute MI, but of those with infarct chest pain, about 59 in every 100 has acute MI.

Let us assume that the serum CPK of this patient has a value greater than 194. 50 percent of patients with acute MI have a serum CPK value greater than 194, with a false positive 20 percent in those without the disease.

P(D) = 0.59 taking the above posterior probability P(ND) = o.41 P(>194D) = 0.50 P (>194/ND) = 0.20

Applying a sequential Bayes’ calculation:

(0.50) (0.59) P (D/>194) = = 0.78 or 78 per cent (0.50) (0.59) + (0.20) (0.41)

Therefore, a patient with infarct chest pain who has a serum CPK less than or equal to 194, the probability of having acute myocardial infarct goes down to 47 percent. As this example shows, any value attached to a finding can be used to modify disease probabilities.

The process of sequential Bayes computation can be carried out for any number of findings. Each time, the posterior probability for one test becomes the prior probability for the next test. Exact probability for diseases and their associated manifestations are required for the Bayesian method as well as for the linear discriminant function. These probabilities must be actually estimated or measured, which is a time consuming and costly activity. Subjective estimations of probabilities by physicians or others have been repeatedly shown to be unsatisfactory and the use of such guess work may result in significant degradation in performance (Leaper, de Dombal, 1972). The future prospects of computerized patient databases will hopefully provide the relevant data used in Bayesian Systems.

Linear Discriminant Function

This represents another statistical pattern classification technique that has been applied to develop CMD systems. Discriminant analysis data from a patient include various patient parameters. For example, consider the question, “What is the risk of neurological complication following carotid angiography in elderly patients with transient strokes?” Faught and associates (1979)

D = 8* Number of transient ischemic attacks; + 6* Number of arteries catheterized +14* Presence (diabetes mellitus ) +11* Presence (sex = female)

They observed that when D was calculated to be more than 55, then the patient had a 77 per cent probability of having a neurological complication from arteriography. When D was less than 55, there was a less than 0.02 percent likelihood of a complication. Thus, the value of D would differentiate high-risk from low-risk patients. A variety of CMD systems have been designed in this fashion, such as for the diagnosis of meningitis (Knall, R., et al., 1977) and lung disease (Matthyas, H., et al., 1979). CMD systems based on the linear discriminant function face the same limitations as discussed under the Bayesian systems.

MULTIVARIATE ANALYSIS

The relations between medical observations (including laboratory data ) are almost always multivariate. It is difficult for the human mind to deal with more than six potentially significant variables in a single judgement. The computer can analyse those multiple analysis has been applied not only to separate patients into disease groups but also to determine which characters or sets of characters will contribute more effectively to the separation. A study was carried out on 15 normal subjects, of which 15 subjects had left ventricular hypertrophy due to hypertension, and 15 had with left ventricular hypertrophy due to aortic incompetence. Together with age, sex, arterial pulse, ballistocardiogram, phonocardiogram, and ECG, the total added up to 45 ‘characters’ for each case. The computer carried out a multivariate analysis, as a result of which it was possible to indicate for each patient a likelihood ratio for the presence of either hypertension or aortic incompetence. In addition, those factors which were most powerful in effecting the separation became evident. The results indicated that one of the hypertensives was classed with the non-hypertensive group and 3 out of 15 with aortic in competence were classed in the hypertensive group. This study is of interest because a considerable degree of segregation of cases was effected without specific medical knowledge, and on the basis of relationships between variables which are not generally recognized clinically. Multivariate analysis of the electrocardiogram has been successfully applied in the computer diagnosis of ischaemic heart disease and myocardial infarction (see also page 151).

DATABASE COMPARISONS AND CASE-BASED RESONING (CBR)

Database comparisons provide a different approach to developing CMD systems centred on the idea of comparing a new patient to similar patients previously treated with known outcomes of treatment. These systems often involve statistical analysis of data, and hence may be profitably described in this section.

Medical knowledge may be explicit, corresponding to the already well-established and domain knowledge, or implicit consisting of individual expertise, organizational practices and past case experience stored in case libraries. Medical case-based Reasoning (CBR) is an effective way of dealing with implicit knowledge (Kolodner, 1993). CBR is analogous to the expert clinician’s analogous reasoning –remembering solutions to similar problems in the past and adapting them to the current situation. Cases actually treated by doctors are the most special source of knowledge apart from medical textbooks, providing the basis for medical case-based reasoning systems (Schmidt, T., and Gierl, L., 2001). The human mind is often biased by the tendency of recalling only more recent cases or more dramatic cases. CBR is a reasoning paradigm able to exploit information embedded into already solved instances of problems called cases. A case consists of three basic pieces of information : (1) the problem description, which is a collection of feature-value pairs able to summarize the problem on hand; (2) the case solution describing the solution adopted for solving the corresponding problem, and (3) case outcome, reporting the results, positive or negative, obtained by applying the solution itself.

The CBR cycle consists of: (1) retrieval of the most similar case (s) from the case library; (2) re –use of their solution to solve the new problem, and (3) evaluation of the outcome.

The development and usage of CBR as a decision-making tool requires a large volume of diagnostically unambiguous clinical cases provided by medical experts (exemplified by the Duke University Cardiovascular Database, described in the next section).

In 1972, Feinstein, et al., demonstrated the usefulness of clinical databases to make prognostic clusters. For example, consider a clinical database on primary lung cancer under the following category.

Oat Cell Carcinoma,

Male Sex, Over 40 years of age, and

Survive for two years or more after diagnosis.

In order to estimate the prognosis of a new case with the same cancer using the database mentioned above, the user first enters the data that describes how many patients already entered into the database and (already treated ) are identical in the cited characteristics of the new patient. The system displays the outcome of those patients. The assumption is that the new patient will have a similar prognosis and a similar therapeutic outcome.

Duke University Cardiovascular Database

The Duk e University Cardiovascular Information System Provides an excellent example of the utility of databases for prognosis and treatment choices. Under the rubric of “ischemic coronary heart disease”, one finds a wide clinical spectrum; at one end is the patient in whom sudden and unexpected death is the first and only symptom; at the other end of the spectrum is a person with stable angina who lives for 80 years. Patients and clinicians would like to know, in advance if possible, who are the patients at high risk for sudden unexpected death or a fatal myocardial infarction, and who are likely to live with their disease for decades. On the basis of items of history, physical examination, stress ECG and other laboratory findings and coronary angiography, the experts at Duke Medical Centres estimated that at least 100 sub-groups would have to be defined to achieve a reasonable likelihood that all members within each group would follow a similar course. Also, considering the number of descriptors per group, the dimensionality goes well beyond the capability of the human brain. Thus, this is the kind of problem that has no solution other than that of computerization.

The Duke Cardiovascular Information System stores the outcomes of the treatment of patients with various sets of attributes. When the attributes of a new patient are entered, the computer selects the most closely matched sub-group. The clinical course and outcomes of all previous patients categorized in the particular sub-group are displayed. The computer’s memory is wholly accurate and dependable, it is unbiased by recent dramatic events and is enhanced by the greater number of entries derived from the entire institution rather than a single clinician’s experience. Therefore, the clinician’s management decision in the case can be based on far more accurate and relevant information than could be possible without the computer.

An interesting feature of this system is that it continuously accumulates and refines its database by means of which increasingly accurate estimates of risks for each intervention can be made. It is important to demonstrate the transportability of conclusions at Duke to other geographic and ethnic groups in different parts of the world.

Atypical variations and rare presentations of common diseases are often missed. A database of a large number of such cases actually treated in the past, provides a rich knowledge base. Pattern-matching techniques allow new cases to be compared with cases in the knowledge base; Rao, N.G. and Guha, S.K., have shown the usefulness of this approach in haematologic problems (MEMCONSULT). Their’s was a pioneer effort in combining CBR with rule-based reasoning –now called multi-modal reasoning.

Clinicians all over the world can make great contributions to future medical managements by helping to create large computerized patient databases for all important diseases. Some examples of databases already created are: on lung cancer (Feinstein, 1972); transient cerebral ischemia (Heyman, et al., 1979); Systemic Lupus Erythematosus (Fries, et al., 1974); general internal medicine (Slameckam et al., 1977). Other related CMD systems use database information to aid the clinician with the diagnostic process (Okada, et al., 1977; Tou, J., 1978).

CBR approach is suitable for peculiar and non-standard situations, with the hypothetico – deductive and rule-based approach is suitable for standard or typical cases. Expert clinician use a multi-modal reasoning (MMR) approach (Montani, S., Bellazzi, R., 2000).

As more computerized patient records (CPRs) become available at websites, web –based Gul provides an interface for users to interact with currently available CBR systems using Java 2.0 programming language with MSSQL. Four prototypes for different purposes, e.g., antibiotic therapy, kidney function diagnosis and time course analysis, are described below as illustrative examples of the CBR approach.

• REFINER (Sharma, S., Sleeman, D., 1988) classifies new cases. • PROTOS, designed for clinical audiology automatically classifies new cases according to previous similar cases. • ICONS (Schmidt, Pollwein, Gierl, 1999) Antiobiotic therapy advice. • COSYL –post operative management of liver transplant patients.

The user has to fill certain details about the case on hand, in a web-based form presented to him. The CBR system will provide the diagnostic support again delivered on the web.

Fuzzy Set Theory

Fuzzy set theory has also has been advocated as applicable to diagnostic problem-solving, (Zaden, l., 1968). The central idea of a fuzzy inference would typically be the assignment of a ‘grade of membership’ in disease classes for a patient, and manipulations of these grades. CMD systems employing the concepts of fuzzy set theory have been developed for several medical problems including renal failure (Kassirer and Gorry, 1978) and assessment of hepatic impairment (Lesmo, et.al., 1980). A fuzzy logic approach has been used for identifying brain structures in MRI images using expert anatomical knowledge (Hillman, G.R. , et.al., 1999).

PRODUCTION RULE SYSTEMS

Al seeks to capture the expertise of humans largely through the use of non-numeric symbol processing. The CMD mimics human decision-making, its reasoning is ‘transparent’ and it has an explanatory capacity. The Al approach most widely used n CMD research is that of rule-based deductions. Medical knowledge is represented as a set of conditional rules or productions. Each rule/production has the basic form, IF/THEN.

IF antecedents THEN consequents-meaning that, if certain antecedents are true, then it logically follows that the consequents are also true. Examples:

Rule Risk 2 IF Sex = Female and age greater than 20 and past sexual history = multiple sexual partners first intercourse before age 18.

THEN cancer at increased risk = cervical cancer Screen 14 IF sex = female and age greater than 20

THEN Recommend pelvic exam annually Screen 17 IF THEN

cancer at increased risk is breast cancer and age from 

20 through 20

THEN Recommend annual breast exam by physician

The inference mechanism in a rule-based CMD system consist of a rule interpreter that applied the rules in the knowledge base of the features of a particular case to reach conclusions about the case. There are a large number of variations on the basic deductive approach. “IF A is true, then B is true; A is known to be true, therefore the logical deduction is that B is also true,” In some systems, rules are applied in an antecedent-driven (bottom-up) fashion where the occurrence of one or more antecedents automatically triggers the application of an appropriate rule to infer its consequents. Conversely, other CMD systems are consequent driven (top-down) where the interpreter, in attempting to establish a certain fact, select a rule with that fact as a consequent and then tries to verify it by confirming that the selected rule’s antecedents are true. In addition, rules can be chained together to make multiple step deductions, e.g., backward chaining or forward chaining (Fig. 10.6) . The selection of goals to be met in backward chaining, or arrived at by forward chaining can be either breadth first in which the goals are achieved all at once, or depth first in which one goal is reached before another is approached. Most expert systems use depth first searchers, with a coherent line of questions being answered by the user (Harman and King, 1985).

For many medical problems, it is possible to reduce the underlying knowledge to a collection of rules where the consequents follow from the antecedents with absolute certainty. In such situations, a measure of the certainty (certainty factor, for example, 0.7) is associated with each rule. These measures can be propagated from one rule to the next as the rules are chained.

Rule-based CMD systems have been developed for a wide range of medical problems in relatively narrow domains, such as the selection of anti-microbial agents for treating infections (MYCIN: Bouchanan and Shortliffe, 1977); diagnosis and treatment of glaucoma (CASNET/GLAUCOMA, Weiss, Kulikowski and Safir, 1978); neurological localization (Reggia, 1978); ventilation monitoring in intensive care units (Fagan, V.M. , et al., 1978); and predictions of drug interactions (Futo,et. al., 1978).

Rule-based deduction systems have provided a symbolic high-level approach to representing problem-specific knowledge in a declarative (that is, non-procedural) fashion. Rule –based systems also support a limited explanatory capability; they can explain their reasoning by citing the chain of rules used to reach a conclusion about a patient. Each rule is a chunk of knowledge and this modularity is claimed to facilitate the incremental acquisition and modification of knowledge bases. Expressing medical problem –solving knowledge as a set of rules can be a fairly difficult task. Knowledge as a collection of rules, since the knowledge acquired through textbooks is in the form of descriptive data.

One must include all the necessary ‘context’ for a rule’s application in its antecedent clauses. If one attempts to incorporate all the relevant contexts into antecedent clauses, there can be a rapid and unmanageable growth in the number of rules present in the knowledge base. It is doubtful whether these limitation will be overcome by adjustments to the underlying production system methodology.

COGNITIVE MODELS

Al experts have tried to mimic the diagnostic approach of the expert clinician –‘hypothesize and test’. Very early in the clinical encounter, an expert clinician builds one or more elementary hypotheses that could explain the patient’s manifestations.

It has been estimated that 75 percent of the verbal questions of the expert clinican are hypotheses –driven, to screen for unexpected manifestation, apart from confirming expected manifestation of diagnostic significance. In the cognitive models of CMD, associative knowledge is represented as ‘frames’ (prototypes, schemata, descriptions), table-like data structures which have been suggested to be a model of human memory organization. Perhaps the best description is that by the creator of the frame systems, Minsky:

“A frame is a data structure in memory for representation of stereotyped situation. We can think of a frame as a network of nodes and relations. Collections of related frames are linked together into frame systems. The frame systems are linked, in turn, by an information retrieval network. When a proposed frame cannot be made to fit reality – when we cannot find terminal as signments that suitably match its terminal market conditions –this network provides a replacement frame. These inter-frame structures facilitate other ways to represent knowledge about facts, analogies and other information useful in ‘understanding’.”

Discussion of an elegant medical programme using the frame system representation of knowledge for acquisition of the present illness can be found in the “Present illness Program” (PIP) produced by Gorry and Parker at the MIT and Tufts University School of Medicine in Boston, USA. It is primarily an understanding oriented system concerned with the explication of clinical cognitive and clinical decision processes. Its goal is the acquisition of a present illness and the formulation of a diagnosis in the domain of renal diseases. The frame-system knowledge representation incorporates a rich descriptive structure and, as a corollary, mandates the incorporation of strategies that can take advantage of this rich descriptive structure.

Each frame represents a hypothesis – a model of a disease or a patho-physiological or clinical state. Separate from the hypothesis, but forming a part of the hypothesis frame, are the findings. Each hypothesis has thus associated with it a set of prototypical findings which may be used to either support or deny the hypothesis. As findings constitute the input by the user, they are matched against the prototypical findings associated with the hypothesis frames via the use of ‘demons’ that watch the traffic in and out of the patient database.

Findings may be of two kinds, delineated in the frames. These two categories are ‘finding’ and ‘trigger’. If the finding input by the user is a trigger for a given hypothesis, that hypothesis frame is activated. If the finding entered by the user is a non-trigger finding, it is ignored by currently inactive hypotheses, but it results in the activation of any currently semi-active hypotheses for which it is a finding. In addition to their role in evoking hypotheses, findings reported by the user are related to the hypothesis frame in terms of logical categorical decision criteria; “is sufficient”, “must have”, or “must not have”. These categorical decision functions participate in building the set of models making up the patient’s specific model. The patient description in thus made up of active or semi-active disease frames and all such frames bear either a complementary or competing relationship to each other. For differentiation of competing hypotheses, a probabilistic scoring function of 4 to 7.

SEMANTIC NETWORKS

As knowledge is concerned with facts and what these facts mean, the semantic network provides a versatile tool for representing knowledge of virtually any type that can be captured in words. Fundamentally, this knowledge is captured in nodes and links. Nodes represent things and links represent meaningful relations between things –things can be persons, objects, concepts, contexts and the like. Semantic relationships may be used to express virtually any relationship that has meaning –causal, temporal, taxonomic, associational, pattern or grouping and the like.

Usually, small number of nodes and links are used to capture isolable elements or chunks of knowledge. These small groupings can be aggregated in logical ways to form higher order networks or models. Hierarchies of nodes and links may thus be established. Such representation is suitable for medical knowledge. Example of semantic networks are already given in Figs. 6.1 to 6.4 in Chapter 6.

Much of the fundamental work in Al has been concerned with searching and traversing the semantic network to obtain efficiency and to avoid a combinational explosion. Although a variety of relationships may be expressed semantically, these may be incorporated into logical groupings and combined, segregated, and other wise manipulated in ways building on the traditions of symbolic logic.

Appropriate combinations may be used to represent models of dysfunction and disease at chosen levels of complexity and resolution. Thus, disease is defined in terms of the patterns of nodes and links that are to be included in the model of a disease or disorder. An illustrative example of a semantic net for lung disease is given in Fig. 10.7.

Specific semantic networks representing a medical knowledge base may use different structure for the aggregation of states and links. The links proposed here are causal links, associated links and grouping links. Causal links should be certain specifications, not only for the type of causality such as ‘may be caused by’, ‘complication of’, etc., and a number representing the likelihood of observing the effect given the cause (a conditional probability), but also specific likelihood numbers that reflect the influence that different aspects of the cause, such as severity and duration, have in leasing to the effect. Also to be considered in such causal features and figures for the influence of contextual features, such as age, sex and weight.

In developing the hierarchical structure of the medical knowledge base, several aggregation processes such as temporal, component, constituent, causal and link aggregations are important. Semantic nets support strategic procedures for confirmation, exclusion, discrimination, planning and further exploration. The latter includes the delineation of the stage and extent of disease and development of prognostic projections essential for the choice of therapy.

Demise of the “Greek Oracle”

INTERMNIST –1 was a pioneering effort in the construction of a diagnostic consultant system for general internal medicine. It was constructed at the University of Pittsburgh by Dr. Jack Myers, an internal medicine expert, in collaboration with Dr. Harry Pople, an expert in computer science, between 1972 and 1982 (Miller, Pople and Myers, 1982). The goal was to create a computer programme that would perform diagnostic reasoning at the level of an expert internist. The style of the diagnostic consultation embodied the ‘Greek Oracle’ model; unable to solve a diagnostic problem, the physician would transfer all the relevant patient information (history items, physical examination findings and laboratory data ) to the computer. during the process, the physician’s role (after the initial data input) was that of a passive observer. At most, the physician could answer ‘Yes’ or ‘No’, to questions posed by the computer program. At the end of the consultation, the program would be expected to reveal the correct diagnosis and to provide a detailed explanation of its reasoning.

There are problems associated with implementing the ‘Greek Oracle’ model in actual clinical practice. The clinician cannot convey his complex understanding to the patient to a computer. further, he cannot afford the 30 to 90 minutes of time required for a complete INTERNIST-1 diagnostic consultation. Dr. Myers, along with Dr. Randolf Miller and Dr. Fred Masarie, developed QMR (Quick Medical Reference) between 1985 and 1986 as a successor to INTERNIST-1. The ‘Greek Oracle’ metaphor is not used in QMR. The clinician’s difficulty in making a diagnosis in a particular case is rarely due to a global failure of the entire diagnostic process, yet this is assumed by the ‘Greek Oracle’ model. The clinician’s diagnostic dilemmas are more likely to occur because of difficulty with one step or a very limited number of steps in diagnostic problem-solving. In QMR, the clinician is allowed to select which steps in diagnostic reasoning require support. QMR acts like a catalyst and requires only partial description of the patient’s state the system is capable of ‘keeping up’ with the physician’s deliberations by providing decision-support tools for selected steps in the physician’s diagnostic reasoning. The ability of the physician user to guide the decision –making process is an extremely important consideration in the eventual acceptance of CMD systems. Table 10.2 gives a sample list of a range of CMD systems. Currently, four expert systems covering a large area of internal medicine are available: QMR, ILIAD, MDX and DXPlain. Al-Rheum is confined to rheumatology. These are described in the next section.

Quick Medical Reference (QMR)

QMR was created following a decade of work on the INTERNIST-1 diagnostic system. QMR has many roles; it serves as an electronic textbook, a low-level consultant and a diagnostic consultant. INTERNIST-1 was written in LISP and assembly language for a large mainframe. QMR is written in Turbo Pascal to run on IBM PC-Ats and compatibles. QMR contains many of the ‘expert’ system characteristics of INTERNIST-1 but employs different data structures. Information about the findings associated with each diagnosis in QMR is stored as ‘bitmaps’. Each disease has a bitmap defining its findings profile. Each bit in the map represents a single manifestation. If the bit is on, this means that manifestations are present in the specified disease; if it is off, it is absent in that disease. Bitmaps are also used to identify differential diagnosis lists and associated disease lists. Since there are about 600 diseases in the knowledge base, and eight bits in a byte, an array of 75 bytes can be used to represent the diseases indicated by each manifestation (each bit position corresponds to a specific disease). In order to determine which disease (among the group indicated by the user) could directly cause the findings, one simply uses a logical AND operation between the ‘disease group’ bitmap and the findings’ ‘differential diagnosis’ bit map; the resultant bitmap consists of the disorders in the group causing the finding directly.

Two other bitmap data structures are associated with each finding in the knowledge base; the ‘by link union’ sets for the findings. These bitmaps are the set of diagnoses that can cause one of the diseases directly associated with the finding, and the set of diagnoses that can be caused by diseases directly associated with the finding, respectively. In order to obtain the indirect (once removed) associations of the user’s chosen finding to the user’s selected group of diagnoses one simply intersects the ‘to link union’ bitmap of the user’s findings with the ‘diagnoses group’s bitmaps and repeats the process for the ‘by links union’ of the finding and the diagnoses group’s bitmaps. The two resulting bitmaps are then combined using a logical OR operation to create a bitmap of all diagnoses that might be related to a direct cause of the finding the ‘indirect’ causes of the finding within the designed group.

The use of bitmaps minimizes space requirements because individual diseases or findings can be represented compactly. They also improve the speed because they permit the execution of logical ANDs and ORs across many bits (findings) simultaneously, using very fast computer operations. QMR running on an IBM PC-AT accomplishes in a few seconds what would take INTERNIST-1 several minutes, on a mainframe. QMR provides two different kinds of information services. It provides information about diseases, their symptoms, and inter-relationships with no reference to any particular patient. It also provides information about the possible causes of findings in specific patients; of course, the latter service necessitates entering data about the patient.

Display of the knowledge base is done in the following ways:

1. Display of disease profiles: This includes a list of findings with evoking strength and frequency for each listed finding, and a list of diseases associated with the disease selected. It takes 8-10 seconds for the computer to accept a diagnosis from the user, and organize and begin to print a disease profile. 2. Display of differential diagnosis: In this case, the user identifies a specific finding, and the computer lists all the diseases that could cause it, in descending order of the finding’s evoking strengths for the listed disease. 3. Display associated of findings to a group of diagnoses: In this case, the user enters a single finding and a single higher level node from the disease hierarchy. This helps the user to understand inter-relationships between the disease and findings in the knowledge base. He can pose a question such as, “How can diabetes mellitus be related to liver disease?” The computer would then show a list of potential mechanisms including hemorchromatosis, or glucagonoma with hepatic metastases.

QMR case analysis mode permits the user to obtain diagnostic lists related to given patient’s history, physical examination and laboratory findings. The user can describe a patient’s case briefly or in elaborate detail, by identifying up to 95 different positive and 95 different negative findings. These can be stored and easily retrieved and new patient information can be added as it becomes available. Case analysis mode offers a number of different analysis options. The user can analyze the case, generate various diagnostic hypotheses, process residual findings after asserting that one or more diagnoses are present, give a critique of the strengths and weaknesses of any of the postulated diagnostic hypotheses, or produce a global hypothesis that explains the patient’s findings on the basis of many inter-related diseases. With the critiquing option the user asks. “What about disease X as a possible diagnosis?” The ensuing critique explains which of the findings are not explained by the diagnosis, and which of the negative findings are strong evidence against the suggested hypothesis. In the critiquing mode, the user can also obtain a list of associated diseases that might explain the findings when the diagnoses itself does not do so.

For complex cases QMR offers two options: ranking hypotheses, and displays possible ‘global case overview. With the first option, the user can view the 15 best diagnoses as determined by the quasi-probabilistic INTERNIST-1 scoring scheme. Hypotheses are ranked on the basis of the positive predictive values of the observed findings for each diagnosis, the ability of each diagnosis to explain the clinically important finding in the case, and the lack of negative evidence contradicting the diagnosis. Thus, a rare disorder that explained all the findings of a patient, normally would not rank as highly as a common disorder explaining all findings except for the least significant ones.

The QMR global case overview option assumes that the patient has multiple inter-related disorders. This program features formulates multiple ‘ground synthesis’ –possible ways of piecing together a complex case like a New England Journal CPC (Clinico-Pathological Conference) case, QMR performs the case analysis functions relatively rapidly.

ILIAD

ILIAD, a Windows- based diagnostic consultant and patient simulator system is an outgrowth of the HELP (Health Evaluation by Logical Processing) system developed by Dr. Homer Warner in Salt Lake City, USA, ILIAD’s knowledge base is created by medical experts whose clinical experience is based on a rich patient database of 500,000 cases in internal medicine. ILIAD incorporates medical research from scientific and clinical literature. The minimum hardware requirement to run ILIAD is a 386 CPU with at least 8 MB RAM and hard disk with 20 MB free space and MS Windows 3.1 or higher. A math co-processor is desirable to achieve an acceptable response time.

The ILIAD knowledge base contains 1,130 diseases in internal medicine and 10,012 disease manifestations (symptoms, signs, lab test results). The knowledge base comprises a set of disease frames consisting of disease descriptions which represent the relationship of each disease to its manifestations (e.g., history, physical examination, laboratory findings). These frames are of two types: (1) Probability (Bayesean) frames containing the manifestations of a disease with their true and false positive rates, along with the general likelihood of the disease in the hospital or out-patient population (Which can be altered by the user according to his geographical location and experience), and (2) Deterministic (Boolean) frames with Boolean IF/THEN logic, which contains a list of findings associated with a disease, the frequency of the finding in the medical service inpatient population, the cost of obtaining the information for each finding that is a test, and a logic statement that is used to determine whether the disease is true.

The knowledge base contains many complex clusters associated with findings because they are the manifestations of some underlying patho-physiological process. Most clusters represent concepts or syndromes with which clinicians are already familiar. A cluster is usually defined using a frame built around a Boolean relationship (IF/THEN rule), clusters can also be Bayesean frames as for example the frame ‘Risk of atherosclerosis’. Clusters may be findings for more than one disease.

Much of ILIAD’s power to mimic the expert physician’s behaviour is based on the ways in which these frames can interact with one another. Information on the status of any frame is passed freely among both types of frames. This provides ILIAD with the ability to make inferences from partial information. ILIAD can be used in consult, critique, test on simulated patient or browse modes. The browse menu contains six options: browse findings, disease profiles, treatment alternatives associated with diseases in the ILIAD knowledge base; ICD codes for diseases in the knowledge base; literature abstracts from the Mosby Year Book of medicine; and viewing pictures. ILIAD’s knowledge base continues to expand and is nearing completion for internal medicine, as a result of intensive knowledge engineering effort in all domains. This system deserves widespread utilization (see also page 312).

Al/RHEUM Consulting System

Al/RHEUM is a knowledge-based consultant system in rheumatology. It is designed to offer assistance in the diagnosis and management of rheumatologic cases to physicians without special training in rheumatologic. The system was developed at the University of Missouri, Columbia, in collaboration with computer scientists at Rutgers, the State University of New Jersey, USA. Al/RHUM was built using a domain independent software tool or ‘shell’ for developing expert systems called EXPERT, produced by the computer scientists, Casimir Kulikowski and Sholom Weiss.

EXPERT is unusual among the Al community on account of its being written in FORTRAN, while most Al systems are written in LISP or PROLOG. The system’s reasoning is accomplished by means of production rules –IF/THEN statements of premises and conclusions. The flow of reasoning for Al/RHEUM moves from its 877 patient findings through 467 intermediate hypotheses to 26 potential disease conclusions. The system will reason with whatever information to trigger any of the disease conclusions.

Rules are categorized into three types: (1) The finding-to-finding (FF)type, used to control branching during the data entry process; (2) The finding-to-hypothesis (FH) type, in which combinations of findings on the left hand side (the THEN part) or (3) the hypothesis –hypothesis (HH) type, in which a context must first be satisfied after which findings and hypothesis on the left side are used to set a confidence level of a single hypothesis on the right.

A disease criteria table Al/RHEUM consists of major and minor decision elements; the required (‘must have’) elements and exclusionary (‘must not have’) elements are combined as specified by the subject matter experts, to trigger conclusions at the ‘definite’, ‘probable’ pr ‘possible’ confidence levels. Al/RHEUM was developed on a DEC system 2060 mainframe computer. Later the EXPERT and Al/RHEUM were moved to the VAX 11/780 and later to a MG 8000 supermicro and IBM PC-AT. It runs under DOS 3.1, requiring the full 640 K of base memory on the AT and 512 K of extended memory beginning at the 1 Mb address boundary. The system can accept nearly 900 findings within about six minutes of data entry time. Most of the entry process consists of responses to menus of right or ten choices. For non-rheumatologist users, a series of definitions are offered online for findings whose meanings may not be fully clear to the user. These ‘tell me more’ items are available to the user by simply pointing and ticking anywhere on the line of the text that represents the findings. The ‘show me more’ and ‘ask MEDLINE’ choices will offer video disc images for findings of a visual nature, as well as direct dial-out from Al/RHEUM to MEDLARS system for a current literature search.

The Al/RHEUM diagnostic output statement consists of a set of disease hypothesis in a differential diagnosis. Each component of the differential is further characterized as definite, probable or possible. Then a summary of the reasoning that generated each component of the differential diagnosis is presented in terms of: (1) the findings that support the diagnosis;(2) findings true for the patient that are not expected, given the diagnosis; and (3) findings currently unknown which, if known and positive, would strengthen the conclusion.

DECISION ANALYSIS IN CLINICAL MEDICINE

Clinicians are constantly required to take decisions under conditions of uncertainty. The process of clinical decision –making has come under scrutiny of experts of formal logical and decision analysis. In 1959, Ledley and Lusted published the first application of formal decision analysis to clinical medicine. Since then, a large body of literature has appeared which has evolved a highly structured system of clinical decision analysis. It has not gained widespread acceptance and has been the subject of much criticism from with in the medical community. But with increasing technology and the widespread use of computers, decision analysis will have an important role to play in medicine in the future.

Formal decision analysis uses the language of probability to replace the uncertainties of clinical data and their relationship to disease. It also allows the decision-maker to attach values to clinical outcomes so that a quantitative solution to the problem can be achieved. It is thus a prescriptive form that tells what should be done in a given situation. It can be applied to individual patients or to whole populations by measuring outcomes in terms of life-years, cost effectiveness, cost-benefit rations, etc. Figure 10.4 depicts the diagnostic and management processes and the place of decision analysis in the management process.

Methodology of Decision Analysis

By considering explicit strategies and modeling their specific risks and benefits, decision analysis is a powerful tool for clinical decision –making under conditions of uncertainty. It integrates published data and identifies critical parameters in the decision –making process. This approach can be applied to many important clinical activities such as evaluation of cancer surveillance programs.

A decision analysis requires the following seven steps:

1. Framing the clinical problem into a focused question; 2. Specifying the possible strategies; 3. Determining the possible risks and benefits associated with each strategy; 4. Quantifying the probability of each of these events; 5. Assigning a value or utility to each outcome; 6. ‘Folding back’ the decision tree to determine the average expected outcome, typically life expectancy, associated with each strategy; and 7. Exploring the effect of uncertainty surrounding the values used, with sensitivity analysis.

Figure 10.8 shows a decision tree with choice nodes and chance nodes.

In sensitivity analysis (not to be confused with the sensitivity of a test), each parameter is varied over a broad range to determine the effects on the results. If the preferred strategy changes during a sensitivity analysis, the model is ‘sensitive’ to that value of the parameter. The threshold is the value of the parameter at which the preferred strategy, and no strategy, lead to the same quality adjusted life expectancy.

Illustrative Example –1

A patient with a history of chronic alcohol ingestion becomes febrile, jaundiced and exhibits marked right upper quadrant tenderness and leucocytosis. His serum bilurubin and alkaline phosphatase are markedly elevated and SGOT and LDH are minimally raised. Oral and IV cholecystography cannot be done because of severity of jaundice. The prothrombin time and platelet count are not precluding liver biopsy.

Choices

• Liver biopsy (assuming that it will always correctly identify alcoholic hepatitis); • Laprotomy (obstructive cholangitis); and • Medical conservative management.

Chances i. Mortality of obstructive cholangitis without surgery = 75 per cent. ii. Mortality of surgery itself = 10 per cent iii. Mortality of alcoholic hepatitis = 25 per cent iv. Mortality of hepatitis if exposed to surgery = 5 0 per cent v. Patients who present with this clinical picture will have: alcoholic hepatitis 90 per cent of the time, and ascending cholangitis 10 per cent of the time. vi. Mortality of liver biopsy is 0.2 per cent

Outcomes 1. Alcoholic hepatitis without surgery gives 75 per cent survival. 2. Alcoholic hepatitis with surgery gives 50 percent survival. 3. Cholangitis with surgery gives 90 per cet survival. 4. Cholangitis without surgery gives 25 per cent survival.

The construction of the decision tree (Fig. 10.9) indicates the outcomes in terms of the probabilities of survival as well as the probabilities of each clinical state or test result as the arms of the chance nodes. Note that if a biopsy is done, the decision for surgery or no surgery awaits the biopsy result exactly as it would in the clinical setting. The sums of the probabilities at any chance node always equal one.

Each outcome is multiplied by the value (probability) of its respective chance node arm, and the sum of the two products about each chance node is obtained (Fig. 10.10). For example, on the top or surgical decision limb, the value for the chance node is equal to (0.5 x 09) + (0.90 x0.1) = 0.54. A value is likewise obtained for medical management (0.70) and liver biopsy (0.76). In this example, liver biopsy is indicated since it has the highest value (0.76)of the three choices. Note that the best method of management does not ensure the best outcome but only the highest probability of a good outcome. This form of analysis is in contrast to the traditional method of decision –making in which the outcome is frequently the only criterion for assessing the method, so that good results may positively reinforce bad methods and bad results may negatively reinforce good methods. It should also be noted that other outcomes could have been used, such as surgical morbidity, biopsy morbidity and so on.

A major criticism of formal decision analysis has been that the input date (probabilities) are frequently not verifiable and the values are subject to doubt. Data based on the results of large and well-designed studies are certainly reliable. In the absence of such data, the use of the ‘best guess’ made from current understanding is open to question. A method for handling this problem is called sensitivity analysis (this is not to be confused with the sensitivity of a test). This method allows substitution of the reasonable maximum and minimum values of the probability in question in to the decision tree. If the conclusion is unchanged, then the probability can be considered as valid. For instance, if one feels that the mortality fro alcoholic hepatitis is more, say 50 percent or 75 per cent, then these values are used in the table and the problem is solved using each as an outcome. If the conclusion of the tree is unchanged, then it is valid for that set of circumstances. Any of the variables (probabilities) may be subjected to sensitivity analysis and if a reasonable maximum or minimum value changes the conclusion of the analysis, then the method should not be considered valid for that clinical problem.

Another consideration that enters the decision process is the question of how much benefit can be expected of a test result and whether the clinical information derived from the test justifies the risk (and cost) involved. In the above example is a mortality of 0.2 per cent did not preclude liver biopsy. In order to find the risk value for the biopsy above which we would be wiser to choose another course, we equate the value of the liver biopsy choice limb to the next most desirable limb, in this case medical management only 0.700), leaving the risk of biopsy as unknown. When the values of the two limbs are equal, then the risk of the biopsy makes either alternative equally attractive (Fig. 10.11).

Thus, if the mortality risk of the liver biopsy is 8.6 per cent or greater, we should choose the other forms of management.

Finally, any test is less than 100 per cent sensitive and specific in classifying disease. If the liver biopsy misses 5 per cent of the alcoholic hepatitis and incorrectly labels another disease as alcoholic hepatitis 2 per cent of the time, then these considerations must be included in the decision tree. The new tree will be slightly more complicated but can accommodate these differences.

Note that the small choice nodes for surgery vs. no surgery have been omitted since it is assumed that surgery will be done only if the biopsy is negative. The value for the limb is 0.998 x 0.742, which still makes liver biopsy the best alternative.

One might ask what is the least sensitivity of the biopsy that would still justify the test. Again equating the biopsy limb to the medical management limb and solving for the unknown sensitivity, we get:

a. 0.90 [ ( 0.75 x sensitivity ) + 0.50 (1-sensitivity)] + 0.1 [(0.02 x 0.25 ) + (0.90 x 0.98)] = 0.700; b. 0.90 (0.25 sensitivity + 0.5) + 0.1 (0.887) = 0.700; c. 0.225 sensitivity = 0.161; d. Sensitivity = 71.6%.

If the liver biopsy has a sensitivity less than 71.6 per cent for detection of alcoholic hepatitis, there is no point in performing the biopsy.

Illustrative Example 2

A 63 –year old housewife with six grown –up children and a 10 –year old history of stable angina pectoris had undergone total hip replacement eight years ago for primary osteoarthrities. Although she had a post-operative pulmonary embolism, the orthopaedic result was excellent and she was pain –free and fully mobile until one year ago. During the last twelve months, she had experienced increasing pain in the her hip on weight bearing, and at the time of consultation, she was largely confined to a wheelchair though she could walk with a cane or crutches for short distances around her house. Eight months back, she suffered an uncomplicated subendocardial anterior myocardial infarction. Although her recovery was uncomplicated, her long-standing stable angina pectoris persisted and frequently limited her ability to get about on crutches. Her orthopedic surgeon reviewed her case, concluded that she had an aseptic loosening of her femoral component and spelled out the risks and benefits of re-operation as follows:

(a) If only the acetablular component required replacement (and the chances of this wee estimated to be only 25 per cent), the probability of a good result (permitting full ambulation again) was 80 per cent. However, that also meant a 20 percent probability of poor result which would confine her to the wheelchair. Moreover, the good or poor results applied to those who survived the surgery under general anesthesia and this patient’s cardiovascular disease suggested a risk of dying in the perioperative period of upto 5 per cent. (b) If, as was more likely (its chances were judged to be 65 per cent), the femoral component required replacement, the probability of a good result was judged to be lower at 60 percent with a 40 per cent chance of poor result. Once again this applied only to survivors of surgery and the increased duration and extent of this more difficult procedure was judged to double the risk of peri-operative death to 10 per cent. (c) If both components required replacement (probability 10 per cent), the chance of a good result feel slightly below half, to 45 percent, which meant that the probability of her being confined to the wheelchair would be 55 per cent if she survived the surgery. The risk of perioperative death for the combined surgical procedure was judged to be 15 percent. (d) Finally, if no surgery was performed, she would either remain the same as she was now (the chance of this was judged to be 20 per cent), or become worse and be permanently confined to her wheelchair (the chance of this outcome was judged to be 80 per cent).

This is a suitable case for applying decision analysis since we have been able to identify all the alternative actions, treatments, clinical changes and outcomes that could occur for the patient in question. On the basis of this information, a decision tree that displays these elements in their proper time sequence is constructed. The probabilities for the various operative risks and benefits are attached to the decision tree (Figs. 10.13 and 10.14).

The points where the tree branches (‘nodes’) are square (‘choice nodes’) when they are under the clinician’s control, and round (‘chance nodes’) when they are not. Note that the probabilities of good and poor results are multiplied by the probability of surviving surgery at chance nodes AS, A3 and A4. The sum of all probabilities at any chance node always equals 1.

(.80 x .95) + (.20 x .95) + (.05) = 76 + .19 + .05 + 1.0 0.202 + 0.414 + 0.050 = .662

Knowledge Engineering Process

Knowledge engineering refers to the restructuring of knowledge in a particular medical speciality into an explicit from ‘understandable’ to a computer. A knowledge engineer is one who works with medical experts, literature, and patient databases, to extract and consolidate facts and estimates into explicit decision modules or frames stored in a knowledge base. A knowledge engineer’s skills are used to determine whether disease findings are dependent or independent, and whether a disease is best defined using a probabilistic or a deterministic model. Typically, knowledge engineering sessions involve domain experts, library specialists and knowledge engineers. Each includes discussion as to whether to include a finding in a disease frame, whether the weights of the findings are appropriate, and if some findings and symptoms should be clustered. Inter –relations among disease findings are identified and recorded in the word and data relation files. In many expert systems, e.g., ILIAD, the diagnostic knowledge base consists of disease descriptions which represent the relationship of each disease to its manifestations (e.g., history, physical examination, laboratory findings). Some of the disease manifestation may be other frames representing intermediate conclusions. A frame is either Bayesian (probabilistic) or Boolean (deterministic). Much of the power of ILAID to mimic the physician’s behaviour is based upon the way these frames can interact with one another. Studies on problem-solving strategies employed by experts have made it apparent that clinical expertise in difficult medical cases is fairly reliant on causal, patho-physiological reasoning. Only those programs relying on such reasoning would be able to cope with the enormous number of ways in which diseases can present, evolve, and interact with each other. Dr. Harry Pople of the University of Pittsburgh, and Drs. Ramesh Patil, Peter Szlovitz and William Schwartz of MIT, Boston, have all emphasized the importance of including causal and patho-physiological reasoning knowledge to improve the performance of expert systems. Inclusion of patho-physiological knowledge in the program provides a powerful way to establish consistent identification of the most reasonable hypothesis but it adds enormously to the computational task. In order to overcome this difficulty, a strategy has been developed that allows reasoning at multiple levels of details in a hierarchial fashion. Although we are 10-20 years away from an ideal medical expert system, we can at least utilize what has been already achieved for improvement in patient care.

CMD via the Internet : PRO Forma and Solo

PRO forma is a formal knowledge representation language that is capable of capturing the logical and procedural content of a guideline, protocol, case pathway, etc., in the form of a set of “tasks” that can be interpreted or “enacted” by a computer. The PRO forma language supports the following four main classes of tasks:

Action: Any utility required for a patient’s care, ranging from a simple injection to a complex surgical procedure. Enquiry: An action whose principal function is to return information from data entered in a form on a screen to a complex computer based imaging function.

Decision: Any kind of medical choice, such as a decision about a patient’s diagnosis, prognosis, level of risk, treatment, need for referral, etc.

Plan: A set or a sequence of tasks carried out over time, to achieve some clinical objective. PRO forma is defined recursively over plans so that an application can include a hierarchical task structure of any complexity. A range of application have been successfully built to support patient care and clinical research using PRO forma.

CAPSULE : A system for advising on routine prescribing.

RAGm: A system designed to support the assessment of risk of familial breast and ovarian cancer in a primary care screening.

ARNO: A pain control system for cancer patients.

Retro Gram: A system for advising on anti-retro viral therapy for HIV/AIDS.

MACRO: A system for running Internet based multi-centre clinical trails ( www.infermed.com).

ERA : To assist GPs in making decisions about whether or not to refer suspected cancer patients for specialist opinion.

Solo: A generic communications infrastructure that permits clinical guidelines driven by a PRO forma enactment engine to be run in distributed environments such as the Internet.

The Solo has three main components, each of which helps to control and validate the communications flow between the www client and server, and the PRO forma enactment engine running the application.

The Integration of PRO forma decisions support and disease management technology (written in Prolog) with Solo, a communication infrastructure (Java servlets, Java wrapper and CORBA technology), support the complete cycle of guideline development and delivery over the www. Together PRO forma and Solo form a powerful tool to support generalized decision- making and process management by computer at the point of need, via mobile PDAs connected remotely to hospital intranets. (Humber, et. al., 2001).