Modeling Knowledge of the Patient in Acid-Base and Electrolyte Disorders

"To imagine a language means to imagine a form of life," according to the philosophy of Ludwig Wittgenstein [27, p. 8e] Conversely. without the appropriate language, that form of life cannot be imagined. Programs that attempt to reason in ways that simulate the thought processes of human beings must use a language of thought that is adequate to the task. Inability to express in a program the facts, hypotheses, strategies, beliefs-the panoply Of notions arising in human reasoning inevitably diminish such programs' acumen.

Our interest here is in programs intended to advise on difficult medical decisions. A generation of such programs, many reported on in this volume, have been designed, implemented, and shown to be capable of expert-level performance on a limited variety of medical problems. Yet it is clear from an examination not of the performance but of the internal organization of these programs that the sophistication of their capacity to represent most of what they deal with is extremely limited. Almost universally, important notions such as the relations between alternative hypotheses, degrees of belief, pathophysiological models of disease, temporal relationships, diagnostic and therapeutic plans, etc., are either missing altogether or else represented by very simple mechanisms that miss much of the subtlety of human reasoning. Apparently the inherent redundancy of information in most medical cases coupled with the relatively great power of the machine to explore many possibilities, to record and maintain its "thoughts," and to apply strategies systematically and without error account for its ability to overcome the seemingly serious disadvantage of its poor representations.

We have begun to develop new representational techniques for many of the concepts underlying medical reasoning in the course of implementing a program for the diagnosis and therapy of electrolyte and acid-base disorders (ABEL). This chapter is intended to convey the richness of concepts that we believe are necessary to allow major improvements in the capabilities Of AIM programs and to lay out a set of ideas we believe will achieve that goal. The program described here is still under development, and although our implementation already has the ability to describe its model of the patient's condition in much better detail than previous programs, extensive work remains in completing the whole diagnostic program based on these ideas.

In this chapter we describe the overall design of ABEL. We support our choice of this domain for investigating the conceptual issues we have introduced and outline the organization of the total program into which the component discussed here fits. We establish a set of desiderata for ABEL (and other programs). Concentrating our attention on diagnosis, we introduce a proposed representation of medical knowledge, discuss structural methods of encoding uncertainty using our representations, describe the operators that manipulate and build the representation, and present some ideas on how diagnostic problem-solving proceeds using these data.

Domain of Expertise

To explore in depth the representations and reasoning strategies needed for thorough medical reasoning, it is useful to choose carefully some specific domain of medical application that poses just the right challenges. We have chosen the domain of acid-base and electrolyte disturbances for the following reasons.

Our concern is not with acid-base and electrolyte disturbances per se. Our basic purpose is to use this domain to determine whether it is now possible to develop an "artificial intelligence" program which is capable of dealing with all aspects of diagnosis and management at a level comparable to that of an expert.

Scope of the Project

The objective of a program in the diagnostic and treatment domain is "the proper management of the patient." That proper management consists of collecting the relevant information about the patient, identifying the disease process(es) responsible for the patient's illness, and prescribing a proper course of action to correct the patient's condition.

One of the complexities of the task is that its subcomponents do not have well defined boundaries, because a patient may be presented to a clinician at different stages of a disorder's evolution and treatment. During the course of management new information about the past history may become necessary as the diagnostic hypotheses evolve. The diagnosis may also depend on information that is presently unavailable (e.g., serum electrolytes that have been drawn but not reported by the laboratory). The disorder itself may evolve through time, providing additional clues to its identity. Further, at times the response to a certain treatment itself may be the best clue to the diagnosis. Therefore, the clinician must choose the next course of action from a large range of alternatives, which can be broadly classified as gathering more information (much of which may turn out to be irrelevant in the current clinical evaluation), ordering further tests (involving possibly expensive time delays and/or clinical costs), waiting for further development, prescribing therapy or some combination of the above.

At every stage of expert consultation, the program must be able to choose between alternative actions available to it with the objective of maximizing the utility of the action to the patient. This can be achieved in a computer program only by developing a system capable of both diagnosis and therapy, i.e., of choosing between the various alternatives available to a physician during patient care. With this in view we are building ABEL, the Acid-Base and Electrolyte Consultant system, which will ultimately address each of the above mentioned aspects of diagnosis and management. Our objective is to separate and modularize different components of the program's knowledge so that we can evaluate our understanding of each and study their interactions. This should also allow us to further experiment with, redesign, and implement any component of the system without having to reimplement the entire program.

The Acid-Base and Electrolyte Consultant system will therefore consist of four major components: (i) the Patient-Specific Model (PSM), (ii) the Global Decision-Making component, (iii) the Diagnostic component, and (iv) the Therapy component. The PSM describes the physician's understanding of the state of the patient at any point during diagnosis and management. This PSM is used as a central data structure with which other components of the system may reason. The global decision-making component is the top level program which has the responsibility of calling the other programs with specific tasks.

In general the global decision program will instruct the diagnostic program to carry out a task such as "take the initial history," or "elaborate a specific diagnosis." The diagnostic component then performs the specified task and reports the results to the main program. It also modifies the PSM to reflect the revised state of understanding of the patient.

Similarly if the global decision-making program calls the therapy-selection program, it attempts to formulate a set of possible therapies for the patient along with a check-list of items that must be tested before any specific therapy can be recommended. It also identifies information that will help discriminate between alternative therapeutic recommendations. This information then is reformulated and sent to the diagnostic program as the next problem to be solved.

The approach of separating global decision-making, diagnostic and therapeutic activities allows us to make explicit the decision-making that goes on in a physician's reasoning. Is further diagnosis necessary? What treatment should be selected? Should we wait before prescribing further treatment? Can we choose some therapeutic action that would also provide diagnostic information, thus making a further effort directed solely at diagnosis unnecessary at this time.

Although the therapeutic capability of the program is ultimately a key element of die proposed system, it is the development of the patient-specific model and the diagnostic subprogram which first claims our attention. Because diagnosis must encompass past evidence of therapy, most of the conceptual problems of therapeutic evaluation occur in a diagnostic program even when it is decoupled from its eventual therapeutic partner. Therefore, we have concentrated first on die patient-description and diagnostic components of the system, and it is these that will be discussed here. Our work on therapy and on the problem of interfacing diagnosis and therapy is still in a rather preliminary state and will not be considered in this paper.

Desiderata

In this section we present some of the characteristics required of our program (indeed, any comprehensive medical management program) if it is ultimately to be useful and effective as an expert consultant.

Making the Proper Diagnosis

The primary responsibility of the program is to make the proper diagnosis. Without fulfilling this criterion, the program offers little possibility of being clinically useful. A criterion for deciding when a satisfactory diagnosis has been reached (and the program should move on to therapy) should weigh the costs of gathering further information in terms of morbidity, discomfort, time and money vs. the benefits of better diagnosis in terms of an improved management plan and a more reliable prognosis. For example, in situations in which the management plan for each of the diagnostic possibilities is the same, further attempts to distinguish between alternatives do not have any utility. Hence, the diagnosis should be considered sufficient. Furthermore, because the program is able to re-evaluate the diagnosis as the disease evolves, it should be able to defer further diagnosis in favor of "creative indecision" [26].

Mode of Interaction

A distinction is often made between two forms of data acquisition in diagnosis: active and passive [7]. The active approach includes the computer's asking a question in order to obtain each new piece of information from the patient. A passive mode is one in which the program is provided with all the information at one point and must make a diagnosis based on this information. The active process suffers from the shortcoming that the physician may be aware of some facts potentially useful to the diagnostic process, but may not be able to communicate them to the program because each new piece of information must be requested by the program. The passive approach avoids this problem but places the responsibility of identifying relevant information on the physician. But, as the user physician is not expected to be an expert in the medical domain of the program, this restriction is unacceptable. Therefore, for the purpose of this program we have chosen a variation of the active process, where the user is allowed to provide any amount of initial information (called presenting information), and from then on the program seeks information in the active mode. At each point in the diagnostic process we allow the user to shift the program into the passive mode so that he can enter new information which he thinks may be useful.

Diagnostic Style

'Me diagnostic style used by the program is almost as important as reaching the correct diagnosis. If the program jumps around among different medical problems under consideration-if its behavior is erratic-the user cannot develop a good model of the program's reasoning. This makes communication between the program and its user confusing and annoying. If the program pursues some low priority diagnostic problem in the face of more important issues, if it ignores a problem of life-threatening character, or if the stream of questions seem pointless, (e.g., if the program continues to ask questions when it should have begun prescribing treatment) it is likely to be rejected by the physician user.

We require that our program exhibit focused, coherent and purposeful behavior in problem-solving, know when to call a halt to its questioning, and be capable of making an interim diagnostic judgment.

Handling Discrepant Information

In virtually any diagnostic workup a large amount of discrepant information must be dealt with. Some of the discrepancies arise because patients are not always accurate observers of their symptoms and because laboratory tests and medical records are often in error. In other cases a seeming discrepancy may arise because of incomplete information, i.e., there may be a valid (but so far unknown) explanation for the apparent disagreement. In addition, there are many cases in which the discrepancy only exists in the context of some currently held but as yet unproven hypothesis. Correct evaluation of each type of discrepancy is critical if the program is to perform effectively. Obviously the expectations of the physician play an important part in identifying and locally evaluating these discrepancies in the incoming information.

Continued Management of the Patient

Typically, a patient is examined by a physician more than once. Thus, the interaction between the patient and physician can be divided into the initial interaction and follow-ups. The follow-up sessions are used by physicians in evaluating their management plans and in refining the diagnosis. In many cases the follow-up sessions are essential for proper diagnosis and appropriate management, and the usefulness of a consulting program is grossly diminished if it cannot follow a given patient over a period of time. Ibis implies a need for the program to maintain a record of its interactions with each patient's case. In the case of the ABEL program, the PSM, which is retained from one session to the next, provides the basis for future interactions.

Explanation

The conclusions reached by the Program should be supportable by a medically meaningful explanation of how the conclusion was reached. Without the ability to explain its behavior and knowledge, the program appears as a "black box"; the theory of its operation is nearly impossible to examine critically or to correct, and its recommendations are likely to be greeted with skepticism by the medical community. Therefore, the communication of reasoning and knowledge used by the program is critical to the credibility and acceptability of the computer program as a clinical aid and also to its usefulness as a model for future development of the technology.

Research in the generation of explanations accessible to its users is not one of the immediate goals of our development Of ABEL, but we have been able to adapt some of the techniques and programs created by Dr. William Swartout for explaining the behavior and recommendations of an expert program for digitalis therapy [24].

Diagnosis

Diagnosis is the process of seeking information and identifying disease processes causing the patient's illness. It consists of ascertaining what the facts are and what the facts mean. Both are essential. Identifying the patient's illness clearly depends on the facts that are already known: less obviously, however, the process of active information gathering depends on the program's understanding and analysis of the available facts. From protocol analysis, researchers have observed that doctors generally use a set of hypotheses (tentative diagnoses) to organize their beliefs and to search for new information efficiently [5, 8, 10, 15, 21]. Therefore, it seems reasonable to organize the program around a problem-solving paradigm using a mechanism for generation and revision of hypotheses. The use of hypotheses allows the program to organize sets of patient-specific findings into a small number of chunks. It also provides the program with a framework to help communicate its "thought processes" to the clinician.

As we argued in the introduction, powerful facilities for representing medical knowledge in much of its innate subtlety are needed. From our experience with existing diagnostic systems and a careful evaluation of the Present Illness Program [15, 16, 25, 26], we are convinced that the power of existing representations is inadequate to state what the physician understands about a patient's illness, and therefore provides an insufficient base on which to build a program that achieves our desired level of expertise. The patient-description must unify all the known facts about the patient, their suspected interrelationships, the hypotheses, and how the hypotheses account for various known and hypothesized findings. As the physician's knowledge is expressed at various levels of detail, from deep physiological and causal knowledge to global knowledge about syndromic and phenomenological correlations, we need a representational scheme that can unify these descriptions. We also need mechanisms for moving from one level of description to another. We have developed a patient-specific model (the PSM) and an associated knowledge-representation scheme which attempts to construct a hierarchic description with the above objectives in mind. This is the central patient-specific data structure of the system.

How is a program to sift through the countless possible hypotheses it could entertain to find those most useful for furthering the investigation and ultimately for solving the diagnostic problem? We suggest that the best scheme is for the program to combine all its evidence into a single expression of its understanding, to use that expression to suggest new data to be gathered, and to revise the expression to incorporate the newly-gained information. We assume that in our medical domain a small set of initial findings will give us the capability of generating a few most promising key hypotheses. Then we employ the hypothesize and reformulate method to revise and refine the program's PSM.

To appreciate the advantages of this approach, consider for the moment an alternative program structure in which after each newly-discovered bit of information the program generates de novo a new set of hypotheses. Stich a process is inherently inefficient because it requires a large amount of information processing (much of it repetitive) and lacks the continuity of thought process so essential to human problem-solving and to the human understanding of program's reasoning,

Uncertainty in our general medical knowledge and in our knowledge of facts about each individual patient creates the diagnostic problem. 'Me task of diagnosis ideally involves confirming an hypothesis (or hypotheses) and eliminating all other competing hypotheses. Multiple possible diagnoses are initially entertained because several disorders may actually coexist in a patient and because the diagnostician may not be able to eliminate all but the right possibility. We note that uncertainty has two distinct components, which should not be treated equivalently. First, lack of precision in our knowledge of disease processes and lack of tools that make perfect observations introduce chance as a source of uncertainty. For example, we cannot predict which of a set of seemingly- identical patients will respond to some therapy and which will not, although we may know that in general a certain fraction respond. Second, our lack of information about a particular patient contributes to uncertainty because before that information is obtained any of numerous alternatives may have to be considered possible. The chief difference between these two components is that chance can only be measured or estimated, whereas gaps in our knowledge of a patient may often be eliminated at the cost and effort of further investigation.⁽²⁾

Uncertainty due to chance is reflected in the patient-description by hypotheses with a low belief-value (or probability, score, certainty factor, etc.) whereas lack of information is reflected by a multiplicity of tenable competing interpretations of the data. The task of the diagnostic problem-solver is then to identify and (if possible) eliminate the uncertainty due to chance and to the lack of information in the patient-description.

The only method available to the problem-solver is to seek new information about the patient. A simple information acquisition strategy would pursue each piece of needed information independently of its need for other data. For example, the program could ask questions intended to explore whether each potential hypothesis is adequately supported. Such a strategy corresponds to trying to confirm all suggested hypotheses. Alternatively, the program could consider in groups each set of hypotheses resulting from a lack of knowledge about the patient and could suggest questions to select one from among each group. This is a version of differential diagnosis.

The principal drawback of the above information acquisition strategy is that it is too unfocused and potentially inefficient. It would attempt to confirm each tenable hypothesis and to differentiate among each set of suggested alternatives independently, without taking into consideration the other acquisition tasks it also faces. Let us call each instance of attempted confirmation and attempted differentiation a problem. Then, instead of pursuing each separately, we propose to collect them all into a single problem-set, and then to apply planning techniques to extract a proposed sequence of questions that most efficiently and systematically reduces uncertainty in each problem being considered. Fortunately this is possible because many of the problems being considered hinge on the same set of observations.

The problem-set will be used to formulate a top-level information-gathering goal. A plan for problem-solving will be generated by decomposing this goal successively into subgoals in the context of the patient-description. Each goal in the plan is associated with the parts of the patient-description relevant to it and with prior expectations about the outcome of the problem-solving effort. The association between the goal structure and the patient-description allows us to separate patient-specific information relevant to the immediate diagnostic problem from information not directly relevant. The expectations associated with the goal structure provide us with the context in which to evaluate the incoming information for discrepancies.

The decomposition of the diagnostic goal into subgoals will be done using various diagnostic strategies based on the protocol analyses by Miller [101, Kassirer [81 and Elstein [5]. The strategies are Confirm, Rule-Out, Differentiate, Group-and-Differentiate, Refine and Explore (the first three strategies have also been used in INTERNIST-I [18], but their application is quite different). The choice of an appropriate strategy for a given situation can be based on the number of hypotheses in the primary goal and their relative belief values.

Representation of Medical Knowledge

The knowledge used by a diagnostic program can be divided into two classes: knowledge about disease and its presentation in a patient, and heuristic problem-solving knowledge. In this section we discuss the representation of knowledge about disease, both general and patient-specific, leaving problem-solving knowledge for later. As the representations of general and of patient-specific knowledge share similar descriptive schemes, we will deal with them concurrently.

Node

Illness can be described as a change in the normal state or function in a patient. To describe an illness, we need a formalism to represent die states, the state changes, die normal and the abnormal functions and their interactions in a patient in terms of the primitives known to the system. It is also important to recognize various composite situations in order to get a global perspective of the patient's illness. Recognition of these situations provides a diagnostic system with tile important ability to reason at a high level of abstraction, organizing a large number of seemingly unrelated facts; more importantly, it may allow the system to use clinical diagnostic knowledge which is organized around high level concepts. Therefore, die diagnostic system should allow descriptions to be in terms of both high level concepts such as diseases and detailed physiological states and processes underlying the presentation of the illness in a patient.

The program's knowledge of medicine and information and hypotheses about thee patient are expressed in terms of nodes that represent states and state changes, and links that express connections between the nodes. Nodes may be primitive if they contain no internal structure or composite if they may be defined in terms of other nodes. A node in the system is described by its temporal characteristics, severity and other aspects relevant to the node. Medically, nodes represent diseases and clinical and pathophysiological states, and normal or abnormal values of quantifiable parameters.

One important function of diagnostic reasoning is to relate causally the diseases and symptoms observed in a patient. These causal relations play a central role in identifying groups of nodes that can be meaningfully aggregated. 'Me presence or absence of a causal relation between a pair of nodes can change their diagnostic, prognostic and etiologic interpretations. Therefore, the system should have the capability of hypothesizing about the presence or absence of a causal link. We also note that, in a manner similar to node descriptions, a causal relation can be described at various levels of abstraction.

Link

A link specifies the relation between two nodes. In the past (pip, INTERNIST, CASNET/Glaucoma), links were used to describe causal relations between nodes. From our study, we have come to the conclusion that this single representation of the interaction between nodes is inadequate, because this representation forces us to assume that every interaction between nodes is causal (or at least statistical) in nature. We believe that the interaction between nodes occurs at various qualitatively distinct levels. For example, two nodes may be causally related to one another, or they may be associated with one another in a statistical sense, without any known causal relation between them, or the presence of one of the nodes may alter the interpretation of the other node without changing the likelihood of the other node in any significant predictable way. To capture these differences, we use the three types of links described below. Further, we believe that a link between a pair of nodes may or may not be present in a given patient (for example, although hypotension and acute tubular necrosis are causally related in general, they may not be related in a given patient). Therefore, in order to reason with relations between a pair of nodes in the patient-specific model, we must instantiate links and incorporate them in the model.⁽³⁾

Causal link. A causal link specifies the "cause-effect" relation between the cause (the antecedent) and the effect (the consequent) nodes. In previous programs, causal links were described by noting die type of causal relation (may-be-caused-by, complication-of, etc.) and a number representing in some form the likelihood (conditional probability) of observing the effect given the cause. However, die conditional probability of observing an effect given the cause (or vice versa) depends upon various aspects of the cause, such as severity, duration etc., as well as other factors in die context in which die link is invoked (such as the age, sex, weight, etc., of die patient and the current hypothesis held by the program). For the effective use of a causal relation, we need to take these conditions into consideration. To illustrate this, let us consider (a simplified) causal relation between diarrhea and metabolic-acidosis. In terms of conditional likelihood, we could state the relation as

(the probability of observing Metabolic-Acidosis given Diarrhea is 0.7). A more complete description of the causal relationship might be given in a rule-like form as follows:

From the above example it is apparent that the conditional probability of observing metabolic-acidosis and its severity and duration depend on the severity and duration of diarrhea and bicarbonate- therapy. Generalizing, it appears reasonable to expect that causal links between the cause and effect nodes should contain the information on how an instance of the cause relates to an instance of effect and on other factors influencing the relation.

To summarize, a causal link in the system is an object denoting the causal relation between a cause-effect pair. It specifies a multivariate relation between various aspects of the cause and the effect, taking into account the context and the assumptions under which the causal relation is being instantiated. A schematic description of a causal link is presented in Fig. 2. The mapping relation permits the computation of a value for any of the attributes of the cause or effect given some (perhaps incomplete) subset of the values of the other attributes. A common use of this facility is to compute some aspect of the description of an effect from knowing its cause; e.g., the severity of the effect from the severity and duration of the cause. The mapping relation may also be used to infer in the opposite direction, for example permitting the calculation of the duration of the cause given its severity and the severity of its effects. Ibis representation of causal relationships and the associated computations permit only purely local information to be used. We have found it necessary to allow the computations on a link to depend in addition on a limited number of non-local factors (the context) and on globally held default assumptions as well.⁽⁴⁾

Associational link. This link indicates that the presence of one node influences the expectation about the presence or absence of the other node. It suggests that the two nodes are correlated, but does not specify the reason for the correlation or association between the nodes. For example (Fig. 3), it is well known that severe hyponatremia is associated with central nervous system disorders such as coma, stupor and confusion. But the specific physiology for the causation of these disorders is not well understood.

Grouping link. Grouping links state that the presence of two nodes simultaneously (in the patient-specific model) represents a situation which is recognized as a part of some useful abstraction. This link does not imply any correlation or causal connection between the nodes connected by the link. In other words this link is used to group together nodes without any commitment to their mutual cause-effect relation. For example, let us consider the group of symptoms: severe hyponatremia, low creatinine (less than 0.6) and normal bicarbonate (Fig. 4). This group of findings is strongly suggestive of die syndrome of inappropriate secretion of antidiuretic hormone (SIADH). They are grouped together in the program, which can then use a recognized instance of the grouping to suggest SIADH at higher levels of aggregation.

Hierarchic Description of Nodes and Links

In this section we discuss a hierarchic representation scheme for the description of the program's patient-specific knowledge. In this representation, the deepest level of the proposed hierarchy describes the knowledge about the patient in terms of primitive physiological concepts and relations known to the system. This physiological description is then successively aggregated into higher level concepts and relations, gradually shifting the emphasis from physiologic to syndromic description or, in other words, from causal to phenomenological descriptions. The phenomenological nature of the aggregate description allows us to use efficiently the hypothesize and match paradigm [111 for global problem-solving. The presence of the aggregated descriptions increases the efficiency of problern-solving by sometimes casting out a large class of possible interpretations of the data. For example, no single disturbance hypothesis about a patient can be consistent with a normal pH. Because the PSM must be causally and physiologically consistent at every level of description, its use restricts the number of hypotheses generated to a small number. We first introduce node and Iink aggregations with the help of examples.

Example 1: Hierarchic description of nodes. To illustrate the concept of a node aggregation, let us consider the condition of excessive loss of lower gastrointestinal fluid (Lower-GI-Loss). The compositions of the lower GI fluid and the plasma are as shown in Fig. 5.

In comparison with plasma, the lower GI fluid is rich in bicarbonate and potassium and is deficient in sodium and chloride (except in the rare disorder characterized by hyperchloremic diarrhea). This information is represented in the knowledge-base by decomposing lower GI fluid into its constituents (and associating appropriate quantitative information with the decomposition) as shown in Fig. 5. The loss of lower GI fluid thus leads to the loss of corresponding quantities of its constituents as shown in Fig. 6 (the quantitative relation is not shown for reasons of clarity).⁽⁵⁾ Therefore, lower GI losses occurring without proper replacement of electrolytes and water will result in hypobicarbonatemia, hypokalemia, hyperchloremia, hypernatremia, and volume-loss, as shown in Fig. 7. At the next level of detail we can describe the causes and consequences of lower GI loss as shown in Fig. 8.

Example 2: Hierarchic description of causal relations. To illustrate the concept of a multi-level description of a causal link, let us consider the following aggregate causal assertion. The causal mechanism of diarrheal dehydration can be explained as follows: diarrhea causes lower gastrointestinal loss which causes dehydration. (Compare Figs. 9 and 10.)

In greater detail, the lower gastrointestinal fluid loss consists of water loss along with the loss of electrolytes (electrolyte losses are not shown here for simplicity of presentation) which results in dehydration if the water loss is not replaced by fluid intake or administration (Fig. 11). Identifying that it is the loss of water that leads to dehydration (as opposed to other components of the GI fluid) is useful because it focuses reasoning on the issues of water balance.

Types of Aggregations

The hierarchic patient-description structure is built by aggregating low level concepts into higher level concepts or by disaggregating high level concepts into their constituents at lower levels. Aggregation allows us to summarize the network of cause and effect nodes describing the patient's illness by substituting aggregate nodes for large chunks of the causal networks and by retaining only the important nodes (analogous to landmarks in a conceptual map [9]). The aggregations and disaggregations can be divided into five classes, described below.

Temporal Aggregation. An observation provides us (clinician or program) with a snapshot of the state of some process. During the course of evolution of a patient's illness, a sequence of snap-shots becomes available to the program. These snap-shots describe the progress of a process through time. Temporal aggregation allows us to resynthesize the process and describe it as a single concept in the patient-specific model. Further, temporal aggregation allows us to summarize a sequence of values of a state variable into a single qualitative description and allows us to make hypotheses about temporal patterns or sequences of values we expect to observe. Fig. 12 shows one example of temporal aggregation.

Component-Summation. Component-summation becomes necessary in our program because the causal links in the system contain specific quantitative information on how the various aspects of cause and effect instances (in a given patient) relate to one another. The matching of cause-effect relations can be unsuccessful if the program is misled into trying to account for some jointly-caused effect in terms of an individual cause. In order to prevent this from happening we must instantiate the effect of each cause individually and then combine them. Different aspects of the components of a compound node may combine additively (if they reinforce each other) or one may offset the other. A component-summation may also be used to compute the magnitude of some aspect of a contributing component from the knowledge of its compound node and its other components.

The description in Fig. 13 is a superimposition of two relations: METABOLIC-ACIDOSIS-1 caused by DIARRHEA-1 and METABOLIC-ACIDOSIS-1 caused by SHOCK-1. The description does not explicitly state that the metabolic-acidosis (in the patient-specific model) is partly caused by diarrhea and partly by shock. Therefore while considering the first relation, the program will assume that the entire metabolic acidosis is being caused by diarrhea. This will result in the cause-effect relation being mismatched (because the severity of acidosis and anion-gap associated with it are larger than can possibly be attributed to diarrhea). A similar mismatch would also occur while considering the relation between the shock and the metabolic-acidosis. To avoid this problem, consider the description of the example as shown in Fig. 14.

The state description of Fig. 13 can be considered as an abstraction of Fig. 14, where METABOLIC-ACIDOSIS-2 and METABOLIC - ACIDOSIS-3 are respectively the components of METABOLIC-ACIDOSIS-1 caused by DIARRHEA-1 and SHOCK-1 The decomposition of an effect with multiple causes into its causal components provides us with valuable information in evaluating prognosis and in formulating therapeutic interventions.

Taking another example, metabolic-acidosis could be considered to be hypobicarbonatemia causing a reduction in pH, which causes hyperventilation and reduced PCO₂ which in turn causes an increase in pH; this is an example of negative feedback. The increase is less than the initial reduction, causing a net reduction in pH. (Fig. 15.)

The component-summation mechanism allows us to separate the primary component of die change from the secondary feedback components in feedback loops and allows us to fold causal chains in feedback loops that represent continuous processes. However, decomposition of a node into its contributing factors raises a new problem, that of how to combine the factors. The combined effect of two components may not be additive and may depend on their causes and on the particular physiological mechanisms involved. This problem needs further in-depth study, but here we assume that there is some local mechanism that allows us to combine the contributing factors satisfactorily.

Node Aggregation. Some of die possible groups of nodes are conceptualized and have objects. In some sense, these concepts provide names describing them as conceptual 0 structures for organizing different situations that are commonly encountered during diagnosis (recall the SIADH example, above). Because of historic evolution in the recognition of medical situations and their nonstandard use, terms such as disease, syndrome, physiological states, etc., have overlapping definitions. But the existence of these different terms signifies the existence of different types of clusters. We will try to identify the basic structure of these clusters without differentiating between the specific mechanisms responsible for their differences, mostly because we do not have sufficient knowledge in the problem-solving domain to exploit the subtle differences between them. An example of causal aggregation is shown in Figs. 13 and 14.

Link Aggregation. Links are used to represent causal relations between nodes. The causal relations are understood and can be described at different degrees of specificity and detail. This has been illustrated in Example 2, above. The link-aggregation hierarchy represents an alternate hierarchy to the node-aggregation hierarchy. It thus provides us with a different viewpoint in summarizing the patient-description by allowing us to identify and eliminate intermediate nodes from causal chains of reasoning. The link aggregation hierarchy also helps us to evaluate the belief in a given causal fink at a higher level of aggregation.

Representation of Uncertainty

In the previous section we have discussed the representation for describing an illness using the assumption that the true state of the patient's illness is known. But in reality the program's knowledge about the patient is uncertain because its knowledge is incomplete and because of the chance nature of disease. Its incompleteness gives rise to more than one possible (hypothesized) diagnosis. Its chance component leads to hypotheses with less than unity likelihoods. In this section we study the two major ways in which this uncertainty appears and extend the patient-description using the concept of hypothesized description to account for the possibility of more than one hypothesized diagnosis.

As we have described above, uncertainty in the patient-description appears in two related but distinct ways. First, because of incompleteness in our knowledge about the true patient-description (ignorance) and second, because of the inherent chance nature of the disease process. To illustrate the differences between the two, let us consider a patient with severe diarrhea and vomiting. This history indicates that the patient almost certainly has some metabolic disturbance, but does not allow us to specify the type of disturbance (which depends upon the relative amounts of alkaline fluid lost due to diarrhea and acidic fluid lost from vomiting). This situation can be represented by specifying a high level of belief for metabolic disturbance and not specifying any belief value for metabolic-acidosis or metabolic-alkalosis. Such a situation prevents the program from drawing any inference about die relative likelihoods of metabolicacidosis and alkalosis from the given information, but permits it to conclude that some metabolic disturbance is present. In general, the existence of a hierarchy of disorders, ranging from very specific states near die bottom to quite general classes of disturbances at the top allows the selection of some particular level in the hierarchy to express the program's proper degree of ignorance. Hypotheses formed near the "bottom" imply a high degree of precision of knowledge, while those at the other extreme imply very little information.

The problem of evaluating the measure of belief in a given hypothesis is a difficult one. Various researchers have worked on this problem [4, 22, 23, 26, 28], but no good practical solution to our problem has been found because the simplifying assumptions often made in using probability theory (e.g., those of completeness of the universe of discourse and conditional independence of events) are not valid here. In addition the issues of incompleteness and uncertainty of information have often been confused. We wish to separate these two issues in order to exploit the capabilities of Al techniques, which are much better at handling incomplete information (by hierarchic description) than dealing with uncertainty. This would allow us to reduce our dependence on the less than adequate schemes available for dealing with uncertainty.

Coherent Hypothesis

At any point in the diagnostic process the program has a partial understanding of the illness of the patient. Uncertainties in that understanding often lead to the program's having several different PSM's, each of which is a consistent account of the observed data. Although each PSM therefore serves as a complete account for everything that is known about the patient, none of them may be an adequate diagnosis. For an hypothesis to count as a diagnostic conclusion, it must not only explain all the known data, but it must also be sufficiently complete that it explains what has caused the patient's current problems and what additional complications are to be expected. It must also be specific enough to differentiate between the actual disease and other similar diseases.

A PSM may fail to be a diagnostic conclusion for several reasons. It may lack any etiologic explanation of the fragmentary state of the patient it describes because no data are yet known about the illness' ultimate causes. The PSM may also lack any predictions concerning expected effects of the current situation, again because no data concerning these effects are yet known and die PSM simply accounts for known data-it does not "speculate." Finally, the PSM may be inadequate as a diagnostic conclusion because it is too non-specific. In the face of very little information, for example, some very general hypothesis such as "congenital heart disease" may be the most specific one justifiable, although additional information would help identify some specific disease that the patient actually suffers from.

We introduce the notion of a coherent hypothesis (CH), which is an extension of the PSM to include guesses about the etiologic derivation of the problem represented by the PSM and about ultimate consequences expected from what is known. Typically, several CH's may be formed from a single PSM, because the known facts are insufficient to determine uniquely the way they came about or their consequences. Each CH provides an alternate explanation for the disease process in the patient and only one of these CH's can be correct. Thus, we have a set of alternate hypotheses which are mutually exclusive and competing; therefore, they can be rank-ordered according to their likelihood.

Alternative explanations for a patient's state may be represented by both the presence of multiple PSM's and the existence of multiple CH's per PSM. Multiple PSM's correspond to different interpretations of already known facts, and multiple CH's associated with a single PSM correspond to different possible complete pictures of the patient's illness that all fit with the given interpretation of the facts known so far. During the diagnostic process, as more facts become known, we expect that both the number of CH's associated with any PSM and the number of PSM's being entertained will decrease as knowledge of the actual state of the patient constrains plausible explanations of the facts.

At any point during the diagnostic process there are only a few distinct explanations for the patient's illness, although each such explanation can have a substantial number of variations. The number of alternatives to consider can be limited by selecting an appropriate level of detail for representing the patient-description. This allows us to represent small differences in the hypotheses implicitly (by professing ignorance about them) while focusing on the major differences. Again, during problem-solving we need to compare the different alternatives (coherent hypotheses) to identify the differences between them. If each CH is represented separately, this task can become substantial. This problem can be overcome by allowing different hypotheses to share common sub-hypothesis structures, thus producing a single structure to represent the set of alternate CH's in which the important differences move up the structure, while the smaller differences tend to be buried deep inside the structure.

Description-Building Operations

In the preceding section we have studied the structural organization of the patient-specific model. Our goal has been to separate the data structures required for the representation of a patient's description from those required for efficient problem-solving. This allows us to represent the patient-description, incorporating all available findings and derived facts about the patient, independently of the program's problem-solving strategies and behavior. The separation improves the program's semantic clarity, coherence, completeness and explainability.

In this section we turn to considering the operations that instantiate the program's general medical knowledge to create descriptive structures pertinent to an individual patient. It is worthwhile noting that the program must distinguish between its general medical knowledge, which is potentially applicable to all patients, and an individual PSM, which describes the program's understanding of a particular patient's case. Any PSM must be constructed from the program's general knowledge and must be consistent with it, for what is known about an individual should not contradict what is known generally. The PSM, however, usually embodies much more specific knowledge about the individual patient than what one would know from a general knowledge of medicine; for example, if a disorder is generally known to have three possible causes, the PSM may in fact identify just one of those as the likely cause in the case under consideration.

The operations for constructing the PSM from the Program's general knowledge and from specific data about the patient are: initial formulation to create an initial patient-description from the presenting complaints, initial findings and lab-results; aggregation to combine various findings into causal clusters representing different disease hypotheses (moving up); disaggregation to decompose aggregate findings and hypotheses into their components or specific subclasses (moving down); and projection to hypothesize associated findings and diseases suggested by nodes in the patient-description at the same level of abstraction (moving sideways).

Initial Formulation

Observations on the response of physicians to the "presenting clinical picture" of a new patient (the initial set of complaints and data) suggest that physicians establish a tentative diagnostic conclusion very early in their thinking about a case. This cognitive strategy has the significant benefit that it enables the physician to organize successive elements of his or her diagnostic behavior (such as what questions to ask, tests to order, observations to make) around the tentative conclusion. This viewpoint provided the organizational framework of the Present Illness Program [15], in which such a tentative diagnosis was established by the triggering of hypotheses strongly suggested by the chief complaint and other important data. Although the triggering scheme was successful in raising the possibility of disorders that should be considered, it was not sufficiently selective [251, and could often trigger as many as half the program's known set of disorders as possibilities.

In the domain of electrolyte and acid-base disorders, the problem of formulating an initial description of the patient from data typically available to the program appears simpler than in the broader diagnostic domain of Pip. ABEL uses any given initial findings and a set of serum electrolyte and acid-base values to construct a small number of PSM's as its initial possible diagnoses, as follows. First it identifies all possible single or multiple acid-base disturbances consistent with the given electrolyte values and then retains only those also consistent with the known initial findings. Next the program generates a possible pathophysiological explanation of the known data based on each of the proposed acid-base disturbances. This is accomplished by using the disaggregation and projection operations (described below) to relate the known clinical information to the known laboratory data. Finally any newly -discovered relationships at the lowest levels of detail are aggregated through intermediate levels to provide their summary at the clinical level. After this process, each PSM contains a plausible formulation of the patient's condition at every level of detail that is of interest to the program. It is these descriptions that are later modified by the diagnostic process as more information becomes available.

Initial formulation of the diagnostic problem could, we believe, be successfully accomplished by a combination of the triggering ideas of Pip with the operators used here to insure consistency across various levels of detail in a causally coherent account of the case. However, the ability to compute the possible sets of acid-base disturbances from a given set of electrolyte values is an especially powerful feature of the domain of study Of ABEL and permits us to avoid some of the issues of initial problem-formulation that would be significant in other domains. We thus follow the dictates of the "knowledge-based systems" approach in Al research to exploit whatever powerful organizing principles or facts of the domain are available. After all, physicians working in the domain also know and rely on these facts to simplify their initial approach to electrolyte and acid-base disorders.

Projection

Within any level of detail of the PSM, we require an operator that explores the diagnostic space. Projection broadens the cross-section of the model at any given level of abstraction. The basic process of projection can be described as follows. Suppose we are considering some hypothesis H. If H is present then some of its antecedents (causes) or consequents (effects) must be present. Therefore, if hypothesis H is assumed, it is reasonable to assume that at least one of its antecedents and many of its consequents will also be present in the patient. The process of projection allows us to suggest new hypotheses about the causes and the consequences of a given node at the same level of abstraction. These new hypotheses can then be used to group different nodes into causally antecedent-consequent pairs, allowing abstraction or disaggregation. Projection is the most powerful operator for suggesting new diagnostic hypotheses within a PSM, because it considers extending the set of already-known nodes to those causally related to them if the new nodes are not inconsistent with what is already known about the patient.

Aggregation

The aggregation operation is used to combine causally related nodes into clusters represented by nodes at a higher level of abstraction. When a set of new findings is entered in the program, they are grouped into causally related clusters about which the program is fairly certain. The program then tries to group these clusters in alternate ways that are consistent with the program's general medical knowledge. If the number of alternate groupings is small (possibly two or three) the program builds alternate structures describing these possibilities. If the number of possibilities is large, the program tries to abstract already formed clusters into semantically larger concepts and tries again. This process is continued until it reaches a level of abstraction where most of the objects are connected to one another or when most of the objects are etiologies or diseases, at which point structural abstraction is no longer useful. Each abstraction generated by this process provides us with a coherent partial diagnosis for the patient, and therefore with a new PSM. Within each PSM, all the diseases, findings, etc., are mutually complementary, while the alternate PSM's provide us with competing diagnoses. In die initial version of ABEL, the program's general knowledge permits only a single grouping of aggregations for any cluster of more detailed nodes: thus, aggregation is used only to summarize detailed information and cannot currently lead to new hypotheses.

Disaggregation

The process of abstraction by aggregation in the system is complemented by the Process of disaggregation. When the program is provided with information about some disease or when it generates some hypothesis at a high level of abstraction, it must assimilate this information into the patient-description. This can only be done if sufficient structure exists at lower levels of the hierarchy to support this hypothesis structurally. If not, the supporting structure must be constructed by decomposing this hypothesis into a more detailed presentation. Here again, we are faced with the situation that there may be many possible disaggregations (presentations) for the same abstract condition. But as this hypothesis must be consistent with already known facts about the patient, parts of the disaggregation will already be present in the patient model and the rest of the disaggregation must be consistent with the model. This reduces the number of possible disaggregations greatly. Even so, the lack of information at lower levels of detail can cause a potentially large number of alternate disaggregations to be possible and thus prevent the program from disaggregating any abstract hypothesis to the lowest level of detail without further discriminating information. Disaggregation proceeds as far as possible without making unsupported assumptions about the lower levels of the PSM.

The mechanisms for creating and manipulating the PSM are described in much more detail and illustrated in a recent paper [13].

Diagnostic Problem-Formulation and Information Gathering

Diagnostic Problem-solving is the acquisition of new information needed to resolve uncertainties in the program's knowledge and the revision of its PSM's to incorporate new findings. The previous section has introduced the set of operators used to manipulate and fie together the different levels of description in the PSM. Here we address die question of the program's method for problem-formulation and information-gathering.

The patient-description assimilates all available information. For a specific diagnostic subproblem, however, a large portion of this information may not be directly relevant. The program must select some subset of its possible interpretations of the facts about a patient, determine what observations would be relevant to reducing its uncertainty about how to interpret those facts, and then take some action to acquire new information relevant to that set. A diagnostic problem-statement should focus on that part of the patient-description where the program's understanding is most uncertain. It should describe alternatives (which can be differentiated), and it should be easily decomposable into smaller subproblems. In this section we discuss die process of identification and formulation of the diagnostic problem, its representation and decomposition into subproblems, the problem-solving strategies, and the use of expectations in identifying discrepant information and in directing the flow of control.

Problem Identification and Formulation

Solving the diagnostic problem involves selecting one interpretation of the information about a patient and assuring that it is sufficiently complete that a diagnostic conclusion and therapeutic and prognostic decisions can be safely based on it. At any point during the process of diagnosis the program has a set of coherent hypotheses (CH's) vying to explain the patient's illness. (Recall that a CH is an extension of a PSM with possible etiologies and ultimate consequences of what is already known.) The specific places where critical information is lacking can be found by identifying points at which, on the basis of the same set of facts, two or more hypotheses seem reasonable. All the diagnostic subproblems identified in this manner are collected into a list called the problem-set.

Every subproblem in this set needs to be solved in order to differentiate between the competing hypotheses. It is here that the information gathering process plays its role of finding critically needed information. The availability of a set of problems to work on simultaneously provides die problem-solver with the ability to minimize the total effort needed in solving all the problems by abstracting common aspects of problems and by selecting an efficient order in which the problems are to be solved. This can be done using either a rank-ordering heuristic or a problem-abstraction heuristic.

From the revised problem-set left after the application of the above heuristics, one problem is selected to be investigated, as described next.

Problem-Description and Goal Structure

Once the problem is formulated, its resolution becomes the top-level goal for diagnostic information-gathering. A diagnostic goal consists of the following components; a) a primary goal which describes the main problem to be solved by the problem-solver, b) a context which describes the reason for solving the problem and c) an expectation which describes the program's prior expectation about the outcome of the problem-solving activity given the knowledge already available to it. The expectations are used to determine whether newly offered information is consistent with the known patient-descriptions. Inconsistencies are used to modify the flow of control in the information gathering program, as described in the section on control flow below. An example of a goal statement is given in Fig. 18.

Once the top level diagnostic goal is identified, the problem-solver sets up a goal structure (a plan for problem-solving) by decomposing this goal into its subgoals (and those, in turn, into their subgoals, in the typical manner of recursive programs) until it reaches subgoals that can be solved using primitives known to the system. The subgoal generation is accomplished using disaggregation and projection operations in the context of the patient-description. Each subgoal is associated with some part of the patient-description relevant to the the problem being decomposed. Therefore, we can view the goal structure as a representation of problem-specific information extracted from the patient-description. This interaction between the goal structure and the PSM allows us to assimilate the information gathered during problem-solving into the patient-description efficiently, because the interpretation and context in which the information is relevant to the patient model is known a priori. It also allows us to associate expectations with goal statements and thus to check the incoming information against the patient-description for apparent and real contradictions.

Problem-solving Strategies

The choice of particular problem-solving strategies is an issue somewhat independent of the representations used to describe what the program knows about a patient's case. Even the above-described structure of problem- formulation allows a variety of methods to be brought to bear on actually solving any one problem (eliminating all but one alternative).

For a long time one type of strategy has dominated the thinking of the medical profession-the differential diagnosis. The codification of this approach in a book such as French's Index of Differential Diagnosis is an example of the systematic organization of such diagnostic procedures. On the other hand, most computer programs using the hypothesize and test paradigm for diagnosis have emphasized the confirmation strategy, in which each hypothesis is considered in turn and some attempt is made to confirm it. Although confirmation is an important strategy in itself, its effectiveness is limited to situations where only one hypothesis is being considered or one hypothesis dominates the competing hypotheses. In ABEL we extend the strategies identified by Miller [10] using our experience with INTERNIST [17, 18, 19] and PIP [15, 16, 25] and adapt them for use in conjunction with the patient-description developed in previous sections.

These strategies can be broadly classified as follows: Confirm, Rule-Out, Differentiate, Group-and-Differentiate, Refine and Explore. Selection of an appropriate strategy is based upon the syntactic structure of the diagnostic problem (e.g., the number of alternate hypotheses being considered and their belief measures relative to one another), as described below. As pointed out above, the confirmation strategy is used when we have only one hypothesis under consideration, or when among a group of hypotheses, one hypothesis is much more likely than all other alternatives under consideration. The rule-out strategy is meant to eliminate some hypothesis, when one is substantially less likely than the others. Differentiation is used to discriminate between two hypotheses with similar belief factors. The group-and-differentiate strategy is used when a large number of alternate hypotheses are held with similar degrees of belief. Here we need to discard many hypotheses rapidly in order to focus our attention on a small number of alternatives. This is done by partitioning the alternatives into a few groups according to some common characterization and then applying a differentiation strategy to rule-out (or confirm) one of the groups, thus narrowing the hypothesis set. The refinement strategy is used to split a hypothesis about a general class of diseases into more specific hypotheses. Note that the refinement of a hypothesis into more specific hypotheses generally results in a disjunctive set of hypotheses. Therefore, the refinement strategy is generally followed by differentiation. Finally, the explore strategy is used when the patient-description does not have any well-defined diagnostic problems to solve. In such a situation we explore the findings systematically, to gather sufficient relevant evidence to formulate a specific diagnostic goal.

An Example

Let us consider a patient who has been ill for 3 to 4 days and is known to have moderately severe metabolic-acidosis and slight hyponatremia (serum Na of 128 meq/L). Let us also assume that no additional history is available. Two possible formulations of the patient's problem are shown in Figs. 16 and 17.

One hypothesis states that the underlying disorder is diarrhea, the other, that it is acute renal failure. The program has set as its top level goal, as shown in Fig. 18, the desire to discriminate between these two possible interpretations.

Goal 1: Differentiate diarrhea, acute renal failure
        Context: Cause-of metabolic acidosis
        Expectations:
            Possible: diarrhea 
                Severity: moderate 
            Possible: acute renal failure
                Severity: moderate
                Duration: a few days
        Subgoals: (AND 2 3 8 9 10)

Fig. 18. Top-level goal structure for discriminating between diarrhea and acute renal failure.

To accomplish this objective the program compares the two interpretations, identifying the differences between the states predicted by the two interpretations, and formulates a set of sub-goals to pursue each difference. For example, the program identifies urinary sodium concentration as a useful differentiator between diarrhea and acute renal failure, because diarrhea predicts that the urinary sodium concentration will be low whereas renal failure predicts a relatively high urinary sodium concentration. Similarly, the program can differentiate between diarrhea and acute renal failure by determining the state of hydration of the patient; this goal can be achieved by confirming either volume depletion or edema. Diarrhea predicts the loss of fluid and therefore volume depletion. Volume depletion, however, can not be directly observed: therefore, the program further decomposes the goal for confirming volume depletion into sub-goals for confirming poor tissue turgor and low blood pressure. Acute renal failure, on the other hand, predicts accumulation of body fluids if normal intake of fluid has continued during the period of oliguria. If a sufficient accumulation occurs, it will manifest itself as edema. A graphic representation of the complete goal structure is shown in Fig.19, and the goals are listed below:

The program has now completed planning its information-gathering strategy for differentiating between diarrhea and acute renal failure. The following is a summary of the information gathered by the program in pursuing this goal structure.

After successfully achieving each of the sub-goals the program evaluates the top-level goal of differentiating between diarrhea and acute renal failure. During this evaluation the program realizes that the overall set of findings are not consistent with one another. In particular, the finding of high urinary sodium concentration suggests acute renal failure, whereas the low serum K concentration is inconsistent with acute renal failure. On the other hand, low serum K concentration is consistent with diarrhea, whereas high urinary sodium concentration is not. This conflict activates the program's excuse mechanism (see below). To resolve the contradiction, the program sets up two goals so that one of the two diagnoses can be confirmed. The two goals are:

Upon pursuing these goals the program finds that the patient has been vomiting, which explains the slightly low serum K concentration and a slight volume depletion. The high urinary sodium excretion argues strongly in favor of acute renal failure.

After completing a cycle of information gathering the program enters this information into the patient-specific models and revises its CH's for the patient and continues the diagnosis.

Control Flow

We have noted that at the time some information is requested, the information gatherer has a hierarchy of goals or a goal stack representing the downward locus of control flow. We have also noted that each subgoal in the goal stack is associated with expectations. The expectations associated with a goal can be viewed as a set of conditions (a predicate) which must be satisfied if the goal is to be attained. 'Me assessment of how well the goal has been achieved is evaluated in terms of one of four values; (i) Success, (ii) Partial Success, (iii) Failure or (iv) Contradiction.

After every question the expectation of the immediate goal is evaluated. If the result is a success, then control is returned to the superior of the immediate goal, with an indication of success. The evaluation results in partial success if the information gathered is not certain enough to satisfy the expectations but can be interpreted in a way that lends support to the immediate goal. On the other hand if the evaluation results in a failure, the problem-solver tries to look for an excuse that can cause the particular goal to be negated without contradiction. This mechanism allows us to distinguish apparent contradiction from real contradiction, providing for a robust process of information acquisition. More importantly, it prevents apparently contradictory information from corrupting the patient-description and thereby compromising the evaluation of hypotheses.

For example, let us consider a patient suffering from the syndrome of inappropriate secretion of antidiuretic hormone (SIADH) who is on a salt-restricted diet due to premenstrual edema of her ankles. Now let us look at the program at a point where SIADH is its leading hypothesis, it has requested information about the urine electrolyte values and is told that the sodium content of the urine is 5 meq/l. Let us also assume that the program is not aware of the salt-restricted diet. If the program now evaluates the SIADH hypothesis, the absence of excessive salt loss in the urine would weigh strongly against the hypothesis and may cause it to drop to a lower rank in the list of hypotheses.⁽⁶⁾ Because SIADH is no longer the program's leading hypothesis, the next question, about a salt restricted diet, which could explain the negative finding, will not be asked. To avoid such situations, the program should delay the scoring of the current hypothesis whenever an unexpected finding is encountered and should explore the finding in greater detail. If an excuse is found, the current goal may result in success or partial success and the program may continue pursuing its goal. However if no excuse is found the pursuit of this goal fails. If the discrepancy is serious, then a mechanism for dealing with contradictions is invoked. In other cases of failure, the expectations of the immediate superior goal are evaluated to determine if the superior goal is still viable or not, and the above-described process is repeated.

If a major discrepancy is found, there is some serious error in the program's patient-description or problem -formulation. This situation is handled with the help of a special contradiction-handler, consisting of two components: a goal structure contradiction-handler and a PSM debugger. The contradiction-handler identifies the scope of the discrepancy. If the contradiction is local to some goal in the problem, the program identifies that goal at which the contradiction does not affect the formulation of superior goals. It then reformulates the problem from there on and continues with information gathering. If on the other hand the contradiction is found in the patient-description, the program must modify the patient-description in light of the new finding and start a new problem-solving cycle. If no reformulation of goals and no revision of the PSM's can resolve the contradiction (a rare event, we presume), then either the recently-introduced fact or some previously-reported fact that it contradicts must be retracted by the user.

The process of back-tracking is computationally expensive. Therefore, commonly occurring contradictions and heuristics for dealing with them Should be pre-compiled in the knowledge-base of the program. Availability of this information will allow the program to resolve commonly occurring contradictory situations efficiently without resorting to explicit back-tracking. The new information gathered during problem solving must continually be added to the patient-description. This process is considerably simplified when the incoming information is consistent with the patient profile and its context is well defined. The use of the goal structure provides a strong focus to the question-asking behavior of the program and provides us with a mechanism to shift from "global" to "local" problem-solving (a shift observed in the behavior of clinicians [21, 10, 81). It also makes explicit the reason for question-asking, and the context in which die question is asked, allowing us to explain why a particular question is asked and the effects of various answers to die question.

Conclusion

in the above sections we have outlined what we believe are a set of needed representations and program mechanisms to enable a "second generation" of AIM programs to improve significantly on the competence of existing programs. We have defined the structure of the ABEL consulting program for acid-base and electrolyte disorders, have introduced the representation with which it builds its understanding of a patient's illness, and have given a sketch of how it can use those structures to plan and carry out its diagnostic reasoning. We have also set forth the arguments that led us to choose this medical domain for our work and summarized a set of desiderata for our program (and others in AIM) to fulfill.

Perhaps the most striking observation as we look back on the years we have already spent developing ABEL is the great complexity of its design and implementation as compared with earlier programs we have developed and examined. To capture the variety and depth of knowledge of expert clinicians requires a major investment. Studies must be made of how physicians reason, and a large body of factual information relating to the specific medical domain must also be gathered. Finally, a set of computational structures must be created to express the accumulated knowledge and to simulate the reasoning strategies. Thus, judging from our experience and that reported by Pople in Chapter 5 of this volume, the size and sophistication of the newer AIM programs will considerably exceed that of the previous generation.

The patient-description component of ABEL has been completely implemented. We are able to perform the initial formulation of the PSM given the basic facts of a case. The description building operators that create and maintain the PSM (aggregation, disaggregation and projection) all function, and we do now have the multi-level causal model of the patient available as the underlying data structure for the program's knowledge. In addition, we have coupled the techniques of English generation described above to the PSM and can now produce reasonably good English explanations of die program's understanding of a case. Figure 20 shows the program's account of the example case used earlier at the final stage where all outstanding questions have been resolved. Compare the concise explanation generated in terms of the clinical, phenomenological level of reasoning in the first part of the figure with the greater level of detail in the second part.

Clinical Level Explanation

This is a 40 year old 70.0 kg male patient. His electrolytes are:

     Na: 140.0            HC03: 15.0            Agap: 13.0
      K: 3.0              pCO2: 30.0
     Cl: 115.0            pH: 7.32

The patient has moderate metabolic acidosis and mild hypokalemia.
The metabolic acidosis causes mild acidemia. The acidemia partly
compensates the suspected moderate hypokalemia leading to the
observed hypokalemia. The metabolic acidosis remains to be
accounted for. The hypokalemia has only been partially accounted
for.

Intermediate Level Explanation

This is a 40 year old 70.0 kg male patient. His electrolytes 
are: ...

The patient has moderate metabolic acidosis, mild hypokalemia and
moderate hypobicarbonatemia. The metabolic acidosis along with 
moderate hypocapnia causes hypobicarbonatemia. The 
hypobicarbonatemia along with hypocapnia causes mild acidemia.
The acidemia partly compensates the suspected moderate hypokalemia
leading to the observed hypokalemia. The metabolic acidosis
remains to be accounted for. The hypokalemia has only been
partially accounted for.

Fig. 20. English language descriptions of the patient-specific model generated by ABEL at two of its five levels of detail for a patient with diarrhea.

Thus far, the implemented diagnostic program is only a simplified approximation to die design presented here. Computer memory limitations have made it difficult to create all coherent hypothesis structures as envisioned in the design, and efforts to combine them into more compact representations have thus far restricted us to using only the explore, confirm and differentiate strategies, and delayed the use of control strategies more complex than simple backtracking. Solving these problems and completing the envisioned diagnosis component is our current focus of effort. As mentioned earlier, the therapy module is only partly designed and awaits further work on die diagnosis module.

How can we put our current status into perspective? We began with a frustration shared by a number of our colleagues after we had built our first programs for medical applications. Although our programs' performance in their domains of expertise was often quite good, even by human standards, we were concerned that we faced fundamental limitations. Our methodology suggests that expert programs should be built to simulate human experts' understanding of and reasoning about problems of the domain. Yet our initial programs clearly lacked machinery to express much of what we knew that our model human experts knew and used. Therefore we set out to envision and build representational structures and reasoning mechanisms that more closely approximated our models. Although completing this work and demonstrating the validity of our underlying assumptions is an extended task whose end is not in sight, the ability Of ABEL to capture and use the kind of rich description of the patient's state that we have shown here is a very encouraging sign.

References

1. Bleich, H. L., "Computer-Based Consultation: Electrolyte and Acid-Base Disorders," Amer. J. Med. 53, (1972), 285.

2. Carlos, V, Pevny, E., and McMath, F., "Computer Analysis of Blood Gases and Acid-Base States," Comp. Biomed. Res. 4, (1971), 623-633.

3. Doyle, J., "A Truth Maintenance System," Artificial Intelligence 12, (1979), 231-272.

4. Duda, R. O., Hart, P. E., and Nilsson, N. J., "Subjective Bayesean methods for rule-based inference systems," Proceedings of the 1976 National Computer Conference, AMPS Press, (June 1976).

5. Elstein, A. S., Shulman, L. A., and Sprafka, S. A., Medical Problem Solving. An Analysis of Clinical Reasoning, Harvard University Press, Cambridge, Mass., (1978).

6. Goldberg, M., Green S., Moss, M. L., Marbach, C. B., and Garfinkel, D., "Computer-based Instruction and Diagnosis of Acid-Base Disorders," Jour. Amer. Med. Assn. 223, (3) (1973), 269-275.

7. Gorry, G. A., and Barnett, G. 0., "Sequential Diagnosis by Computer," Jour. Amer. Med. Assn. 205, (12) (1968), 141-146.

8. Kassirer, J. P., and Gorry, G. A., "Clinical Problem Solving: A Behavioral Analysis," Ann. Int. Med. 89, (1978), 245.

9. Kuipers, B. J., Representing Knowledge of Large-Scale Space, AI TR-418, MIT Artificial Intelligence Lab., Cambridge, Mass., (1978).

10. Miller, P. B., Strategy Selection in Medical Diagnosis; Technical Report TR-153, Project MAC, Massachusetts Institute of Technology. (September 1975).

11. Newell, A., and Simon, H. A., Human Problem Solving, Prentice Hall, (1972).

12. Patil, R. S., Design of a Program for Expert Diagnosis of Acid Base and Electrolyte Disturbances, TM-132, MIT Lab. for Comp. Sci., Cambridge, Mass., (May, 1979).

13. Patil, R. S., Szolovits, P. and Schwartz, W. B., "Causal Understanding of Patient Illness in Medical Diagnosis," Proc. Seventh Intl. Joint Conf on Artificial Intelligence, Yale University, New Haven, CT, (1981).

14. Patil, R. S., Causal Representation of Patient Illness for Electrolyte and Acid-Base Diagnosis, Ph.D. thesis, Dept. of Electrical Engineering and Computer Science, MIT Lab. for Comp. Sci., Cambridge, Mass., (Oct. 1981).

15. Pauker, S. G., Gorry, G. A., Kassirer, J. P., and Schwartz, W. B., "Toward the Simulation of Clinical Cognition: Taking a Present Illness by Computer," Amer. J. Med. 60, (June 1976), 981-995.

16. Pauker, S. G., and Szolovits. P.. "Analyzing and Simulating Taking the History of the Present Illness: Context Formation," in Schneider/Sagvall Hein (Eds.), Computational Linguistics in Medicine, North-Holland, (1977), 109-118.

17. Pople, H. E., Jr., "Artificial-Intelligence Approaches to Computer-Based Medical Consultation," IEEE Intercon Conference, (1975).

18. Pople, H. E.. Jr., Myers, J. D., and Miller, R. A., "The DIALOG Model of Diagnostic Logic and Its Use in Internal Medicine," Advance Papers Fourth Intl Joint Conf. on Artificial Intelligence, MIT Artificial Intelligence Lab., Cambridge, Mass., (1975).

19. Pople, H. E., Jr.. "The Formation of Composite Hypotheses in Diagnostic Problem Solving: an Exercise in Synthetic Reasoning," Proc. Fifth Intl. Joint Conf on Artificial Intelligence, Carnegie-Mellon U., Pittsburgh, Pa., (1977).

20. Richards, B., and Goh, A. E. S., "Computer Assistance in the Treatment of Patients with Acid-Base and Electrolyte Disturbances," in Shires/Wolf, (Eds.), MEDINFO 77, IFIP, NorthHolland Publ. Co., (1977).

21. Rubin, A. D., Hypothesis Formation and Evaluation in Medical Diagnosis, S.M. thesis, Technical Report Al-TR-316, MIT Artificial Intelligence Lab., Cambridge, Mass., (January 1975).

22. Shafer, G., A Mathematical Theory of Evidence, Princeton University Press, (1976),

23. Shortliffe, E. H., and Buchanan, B. G., "A Model of inexact reasoning in medicine," Mathematical Biosciences 23, (1975), 351-379.

24. Swartout, W., Producing Explanations and Justifications of Expert Consulting Programs, Ph.D. thesis, Dept. of Electrical Engineering and Computer Science, TR-251, MIT Lab. for Comp. Sci., Cambridge, Ma., (January 1981).

25. Szolovits, P., and Pauker. S. G., "Research on a Medical Consultation System for Taking the Present Illness," Proceedings of the Third Illinois Conference on Medical Information Systems, University of Illinois at Chicago Circle, (November 1976).

26. Szolovits. P., Pauker, S. G., "Categorical and Probabilistic Reasoning in Medical Diagnosis," Artificial Intelligence 11, (1978), 115-144.

27. Wittgenstein, L., Philosophical Investigations, MacMillan, New York. (1953).

Notes

(1) The research reported here has been supported in part by the National Institutes of Health under Grant No. 1 P01 LM 03374 from the National Library of Medicine and Grant No. 1 P41 RR 01096 front the Division of Research Resources. Dr. Schwartz's research is also supported by the Robert wood Johnson Foundation. Princeton, NJ The program described here is being constructed as part of Mr. Ramesh Patil's doctoral research, with the guidance and assistance of the other authors. Much of the chapter is drawn from Mr. Patil's thesis proposal, which has appeared as a technical memo [12]. This material is also discussed in much greater detail in that thesis [14].

(2) Some may object that the lack of general medical knowledge that leads to a theory admitting chance could also be ameliorated, for example by a research project elucidating the biochemical nature of the therapy under consideration. For our purposes, when considering the diagnosis and therapy of individual cases, this luxury is precluded and the distinction between chance and lack of information remains useful.

(3) By instantiate we mean to create a concrete instance of a general concept to represent that general concept in the particular hypothesis at band.

(4) In computation, a default is a value one may assume in the absence of evidence to the contrary. It is a common rule of thumb, for example, that in hearing a description of a patient one may assume that those aspects of the patient that are not specifically mentioned are normal. Reasoning based on such default assumptions may have to be retracted if real evidence contradicts the default assumptions. [3]

(5) The following abbreviations are used: CONST-OF (constituent of) and COMP-OF (component of).

(6) This reasoning is based on the program's assumption that the patient is on a normal-salt diet, which is reasonable in the absence of knowledge to the contrary. The program must make many such assumptions when working with partial information.

Modeling Knowledge of the Patient in Acid-Base and Electrolyte Disorders(1)

Introduction

Domain of Expertise

Scope of the Project

Fig. 1. The ABEL System is organized around knowledge of the Patient and the program's hypotheses as represented by the Patient-Specific Model. The global decision-making module invokes specialists for diagnosis and for planning and execution of treatment.

Desiderata

Making the Proper Diagnosis

Mode of Interaction

Diagnostic Style

Handling Discrepant Information

Continued Management of the Patient

Explanation

Diagnosis

Representation of Medical Knowledge

Node

Link

Fig. 2 A causal link in ABEL represents not only a cause and effect but a multivariate relationship between various aspects of the cause and effect, and permits incorporation of the effects of the diagnostic context and assumptions.

Fig. 3. Associational links specify a known co-occurrence without causal connection.

Fig 4. Grouping links suggest constellations of findings whose co-occurrence should routinely be recognized. in this instance, the grouping is strongly suggestive of the syndrome of inappropriate secretion of antidiuretic hormone (SIADH).

Hierarchic Description of Nodes and Links

Fig. 5. Composition of lower gastro-intestinal fluid and of plasma.

Fig. 6. Loss of lower gastro-intestinal fluid is accompanied by loss of the electrolytes it contains.

Fig. 7. Direct results of the loss of lower gastro-intestinal fluids.

Fig. 8. Causal relations surrounding lower gastro-intestinal fluid loss at a more aggregate clinical level.

Fig. 9. Clinically, diarrhea is known as a possible cause of dehydration.

Fig. 10. Diarrhea causes dehydration by causing the loss of lower GI fluid.

Fig. 11. The link of Fig. 10, expanded to show the detail of water balance.