Towards a generalized model of diagnostic behaviour

ARCHITECTURE The relationships within the domain of findings and the relationships within the domain of hypotheses are far more numerous and complex than their interrelationships. However, during the initial stages of diagnostic activity, the shifts of reasoning from the findings base to the diagnostic (hypotheses) base are crucial, as they establish the initial context for the problemsolving activity. Clancey is refers to such initial reasoning jumps as 'heuristic inferences' a non-hierarchical and non-definitional connection between concepts of distinct classes. Although such inferences are very nearly categorical, as their name implies they are not infallible. Examples are the constrictor associations in CADUCEUS 19 and the trigger associations in PIP 2°, PUFF el and NEOMYCIN 22. The model of diagnostic behaviour proposed here has an abstract architecture that separates the two bodies of factual knowledge along the lines of conceptual cleavage in the diagnostic concepts (see Figure 2). At this level of analysis the model resembles that proposed by Chandrasekaran and Mitta123. However, at more refined levels of analysis and on more concrete issues, the two models differ significantly. The findings reasoner operates on the conceptual organization of the general findings knowledge to make intelligent inferences on the available case-specific information. Such inferencing could be of a commonsense nature, deducing that a child is a non-smoker, for example, or it could be based on specialist knowledge, e.g. deriving suitable qualitative abstractions from the quantitative results of laboratory tests entails knowledge of the units of measurement used. The diagnostician generates and refines hypotheses by operating on the conceptual organization of the general hypotheses knowledge. A critical function of the diagnostician is in assessing what additional information would be needed for the diagnostic activity to progress. The diagnostic Set of loosely bound More rigid auxi l iary tasks structure FINDINGS ] Heuristic jumps ]DIAGNOSTICIAN R E A S O N E R ] • abstraction J ~ • / • abduction • restriction 1.~ L/uerles | • deduction "defau l ts etc / Replies ] " induction User volunteered information Figure 2. Abstract architecture overview, representing the separation of the findings knowledge and the diagnostic knowledge. The reasoning processes within each of these components and the communication protocols between them are not represented picture is the data structure that holds the user-supplied case-specific information and the results of the operations of the findings reasoner and diagnostician. The diagnostic picture for a particular domain could be of considerable structural complexity. As this data structure essentially holds the instantiations of the general knowledge that apply in the particular case, its complexity would depend on the complexity of the organization of the knowledge bases. The communication between the findings reasoner and the diagnostician could be restricted to message passing through the diagnostic picture, but this is an implementation rather than a conceptual issue. The separation of the data handling function from the main diagnostic function is seen as the first step towards the explication of the total functionality of a diagnostic task. An important feature of the proposed architecture is the distribution of the overall reasoning among a number of knowledge-intensive tasks (see above). The top level distribution is between the diagnostic and data handling tasks. The findings reasoner comprises a set of loosely bound auxiliary tasks (see below), while the tasks comprising the diagnostician are rigidly bound together into a cooperating whole. The distribution of reasoning as such enhances the modularity of the overall system, and by implication, its extensibility. Furthermore, and importantly, this allows the construction of a knowledge engineering environment with acceptable skeleton system tools. REASONING WITHIN A HYPOTHESIS BASE A diagnostic task can be characterized by its breadth (very narrow, e.g. dealing with one disease; or very broad, e.g. dealing with internal medicine) and its depth (very shallow, e.g. reasons from the clinical knowledge; or very deep, e.g. reasons from the Vol 2 No 3 September 1989 167 pathophysiological knowledge). Although at a high level of abstraction every diagnostic process can be captured within a uniform framework (see below and Reference 24), at less abstract levels individual processes can diverge significantly in the way these high level forms of inference are carried out. This is reflected in the wide spectrum of representation structures that are encountered in the knowledge bases of diagnostic systems. There are three forms of diagnostic inference:5: abduction, the process of generating hypotheses; deduction, the process of testing hypotheses; and induction, the process of evaluating hypotheses. Figure 3 qualifies diagnostic steps from this perspective. A hypothesis could be a simple assertion, a collection of property value associations, possibly with certainty measures, an associative network of simpler hypotheses, or a combination of these. The same hypothesis could be described at different levels of abstraction and within the same system: hypotheses could be of different types, e.g. in a medical system some hypotheses are primary etiologies, others are syndromes or intermediate states, etc. By definition, hypotheses can not be directly established through observation, and thus they need to be inferred or gradually pieced together from observations and inferences. Diagnostic inquiries Abductive diagnostic steps Abductive diagnostic steps generate hypotheses of two types: non-contextual and contextual: non-contextual: (findings) -) (hypothesis) contextuah (hypothesis) * (findings) ) ( hypothesis ) contextual: ( hypothesis ) -) ( hypothesis ) Non-contextual abductive steps are Clancey's (simplest) heuristic jumps. They function to set up the initial context for the problem-solving activity. In the proposed model, initial context formation is a collaborative activity between the findings reasoner and the diagnostician (see below). The hypotheses thus generated are usually few and general. Non-contextual abductive steps, therefore, function to significantly constrain the range of possible explanations of the malfunctioning behaviour. A contextual abductive step occurs when a hypothesis is generated in the context defined by another hypothesis (the two hypotheses could be of different types, thus having another case of heuristic jumps). Such steps are qualified either as traversal of a taxonomy or as lateral shifts. In the former case, the generated hypothesis is either a refinement, or a generalization, to the initial given hypothesis, and in the latter case, an opponent, or a complement, to this hypothesis. Generalization shifts broaden the range of currently entertained explanations, and refinement shifts constrain this range even further. Thus, taxonomic shifts of the former type are more likely to occur at the initial stages of the process, and of the latter type at subsequent stages. Opposing hypotheses share expectations, giving rise to situations where the actual presence of the one may be confused for the presence of the NonGiven ( firldings/ ~ conte×tual ] Result <hypothesis • <hypothesis) ~Abductive ! (;iven ] <findings) [or <constraint)) Lc°ntextual " Taxonomic <generalized hyp) ' Result ~-shift -~<ref ined hyp) L Latera[ _ r-<opposing hyp) shift L <complementary hyp) J Diagnostic~ Deductive { Given-(hypothesis)r Given step Result-<expectations) J " <conclusions) <all findings) -Termination l f [-<satisfactory explanation) i i Result <conclusions too general) ~ _ < unex plai ned findings) _ i inductive i <treatment not possible) i <hypothesis) !'Given ] <all unexplained findings) -Other ! F<accept hypothesis) : Result-{ <reject hypothesis) L<update belief in hypothesis) Figure 3. Qualifying diagnostic" steps. { ." joint selection; [ : exclusive selection; ( : selection may be repeated other. The given findings in such abductive steps are not expected on the currently entertained hypothesis, but on an opponent of it. Complementary hypotheses are related via a causality, or a complication-of, or some other sort of an association relationship. Thus, the given findings point to the direction of the particular complementary hypothesis, the objective being that the two hypotheses together wilt give a more accurate and complete picture of what is causing the malfunctioning than either one would do on its own. Groups of complementary hypotheses constitute complex or composite hypotheses. Contextual abductive steps function to piece together the components of more complex and more diagnostically complete (global) hypotheses. Contextual abductive steps do not necessarily involve findings: a hypothesis can directly point to a complementary hypothesis. Contextual abductive diagnostic steps can take place within the scope of inductive diagnostic steps (see below). For example, consider an inductive step concerned with matching a hypothesis' expectations against the reported findings to decide whether the hypothesis can be concluded. If an expectation violation is thus detected, which points in the direction of an opposing hypothesis, a contextual abductive step has occurred. Deductive diagnostic steps Given a hypothesis, deductive diagnostic steps decide which findings follow by necessity from the hypothesis. Thus, actions for testing the unobserved expectations, on the hypothesis, can be requested. In the proposed model the acquisition of new information is again a collaborative activity between the diagnostician and the findings reasoner (see below). Inductive diagnostic steps An inductive inference is from hypotheses to overall best explanation, or termination of the diagnostic process. Usually a satisfactory explanation is when the critical abnormal findings are accounted for by the concluded hypotheses and, more importantly, when the explanation permits the undertaking of ' treatment' procedures. It is only after treatment procedures have 168 Knowledge-Based Systems been undertaken that we can really test the correctness of the induced explanation of the problem. For example, in the case of hardware devices, a satisfactory explanation of a problem would indicate a set of replaceable/repairable units as being faulty. If after replacing/repairing the particular units the problem still exists, then the given explanation was incorrect. Of course, if each contending hypothesis gives rise to the same treatment there is no justification in trying to resolve these hypotheses any further. All three forms of inference should be present in a diagnostic system. We found that evaluating existing systems from this abstract but complete model of diagnostic inference gives rise to many useful insights. It seems that of the three forms of inference, deduction is the one understood best, and induction is the one understood least. Although diagnostic systems have been built that only employ deduction (e.g. MYCINZ6), neither abduction nor induction can suffice on their own. Knowledge engineers agree that inductive processes are the ones most difficult to analyse, How expert diagnosticians evaluate hypotheses, and how they decide when to stop are still largely unanswered questions. The PIP and INTERNIST-127 systems have very ad hoc termination criteria, and the designers of ABEL 2s accept this for their system. Focusing and information acquisition The focusing heuristics are inherently inductive, and the information acquisition heuristics are inherently deductive in nature. However, focusing and information acquisition are intimately related (and critical) aspects of a diagnostic process. The former decide on the part of the hypothesis space on which to focus next, and the fatter decide on the information which is required next. Figure 4 gives a complete abstraction of a diagnostic process from this perspective. To be complete, the analysis of a particular diagnostic task must therefore reveal the heuristics for the initial context formation, the focusing heuristics and the information acquisition heuristics. MYCIN has a simple parameter-value language. The system reasons in a backward-chaining fashion, as such goals are always more general than the premises. The whole notion of hypotheses is alien to MYCIN (see Clancey29). At the heart of a diagnostic process' focusing aspect lie the means for evaluating the promise of hypotheses. As mentioned above, eliciting these means for a particular domain could prove to be a very challenging undertaking. Existing diagnostic systems have been criticized for erratic changes in focus which are not akin to the corresponding human reasoning. As focusing and information acquisition are so intimately related, problems with the one aspect in a particular system may m fact be related to errors in the other aspect, i.e. focusing errors can propagate themselves as information acquisition problems, and vice versa. This fact should be borne in mind during the evaluation stages of a diagnostic system. For example, MYCIN's information acquisition problems are due to the lack of focusing, and PIP's focusing problems are to a large extent dependent on the system's shortcomings in its information acquisition. ~ / ~ Focusing heuristics Terminate I nductive A b d u c t i v e In i t ia l con tex t formation heur is t i cs I "t'n'°rma"°°l acquisition ] : ,,, heur is t i cs