This section is part of
Patil, Ramesh S. Causal Representation of Patient Illness for Electrolyte and Acid-Base Diagnosis. MIT Lab for Comp. Sci. TR-267 (1981).

7. Conclusion

7.1 Summary

Each new scientific endeavor is built on previous endeavors, consolidating their successes and learning from their shortcomings. This is no exception; we have drawn heavily from first generation AIM programs. This thesis has benefited from the studies of clinical skills of physicians, by introspection and by observing each other, development of models of diagnostic processes and their implementation using computers by Schwartz, Pauker, Gorry, Kassirer, Szolovits and others. Implementation of the Present Illness Program and analysis of its performance was an important first step for the research presented here. Experience with PIP and the other first generation AIM programs exposed the need for substantially more detailed and categorical reasoning in diagnostic programs and provided an ideal environment in which to explore the issues addressed in this thesis.

The research presented in this thesis was also influenced by the discussions of difficult diagnostic cases at the informal Electrolyte and Acid-Base rounds at the Tufts New England Medical Center Hospital. The most striking aspect of these discussions was the frequent use, by the clinicians, of the pathophysiological knowledge in evaluation and justification of diagnostic hypotheses, and the ease with which they were able to combine knowledge of global diagnostic associations such as “disease X is a common complication in a patient with a history of Y” with intricate physiological deductions such as “Na+ and K+ exchange in the distal tubule is coupled with the excretion of H+, therefore increased distal delivery of Na+ enhances ...” These observations strengthened our conviction that in order to begin to approach the level of competence of an expert a computer program must possess a similar ability. It must be able to reason simultaneously with phenomenological knowledge about disease associations and with the best available pathophysiological knowledge about disease mechanisms. Much of our effort has been focused on building representational and procedural mechanisms to provide such a capability. The emphasis has been on the development of multi-level causal descriptions of a patient's illness and on the development of techniques for composing/decomposing effects with multiple causes (described in chapters 3 and 4). We believe this approach provides our program with a level of understanding of disease not present in the first generation of AIM programs.

The study of clinical problem solving activity by Elstein [Elstein78], Kassirer [Kassirer78] and others suggests that a physician's diagnostic reasoning is strongly guided by structural notions such as “coherence” and “adequacy”. Each diagnostic alternative entertained by a physician is a mosaic of connected hypotheses, together accounting for the observable aspects of the patients illness. This thesis describes the use of a coherent hypothesis as the logical unit of hypothesis representation (represented as a PSM). A PSM is a collection of causally connected disease hypotheses and findings, providing a (perhaps partial) explanation of the patient's illness. A set of alternative diagnoses consistent with a PSM is represented using a diagnostic closure. A diagnostic closure unites all the dependencies and expectations necessary for diagnostic inquiry, for selecting appropriate questions, and for evaluating the information received in response to the questions.

Expert clinicians employ a variety of diagnostic strategies for an efficient exploration of the diagnostic space. Some of the first generation programs, notably INTERNIST-I, use similar strategies to guide their diagnostic exploration. However, their lack of commitment in pursuing a given strategy to completion results in unfocused and inefficient use of these strategies. This problem can be alleviated by allowing the diagnostic problem solver to plan a sequence of questions focused around a diagnostic task before embarking on an inquiry. In this thesis we have described a simple diagnostic planner which formulates a tree structured plan. The planning begins with the global task of discriminating among the set of alternative PSMs (and their associated diagnostic closures). This task is reduced to a set of questions by recursive application of diagnostic strategies: confirm, differentiate, rule-out, group-and-differentiate and explore. The diagnostic planning provides the program with a focused and efficient diagnostic behavior. In addition it serves as a framework for justifying the motivation for asking a particular question.

We have argued that for a competent medical system to be accepted, it must be able to explain its conclusions to its user. This thesis has applied some recent explanation technology [Swartout80]35 developed in a simpler domain to the much more complex domain of electrolyte and acid-base diagnosis. ABEL is capable of justifying its motivation for asking a particular question and explaining its understanding of the patient's illness at multiple levels of detail.

7.2 Limitations of ABEL and Future Directions

The research presented in this thesis has many limitations. Some are due to limitations of time and resources. More seriously, the inherent size and complexity of the domain has forced us to limit the scope of this research to just a few issues and to adopt engineering compromises.

The representation of the relation between states in ABEL is inadequate; all interactions are described using a single type of link, namely a causal link. This is unnatural when the relationship between disease states is statistical with no known causal explanation. Furthermore, we need to group states which jointly have significant diagnostic and prognostic implications even if the states are not causally or statistically related. Weaker relations, such as “associational links” and “grouping links” are needed to capture these two cases [Pauker76, Patil79].

Causal interactions are themselves complex and multi-faceted. For example, an effect may be triggered by a cause, or the presence of an effect may require continuous presence of the cause. We consider an elaborate taxonomy of causal relations (e.g., [Rieger77]) to be a necessary component in the future development of ABEL.

One primary objective of this thesis has been to explore structural criteria such as coherence and adequacy in the construction and evaluation of causal hypotheses. We have intentionally avoided probabilistic measures in order to test the full potential of this newly developed structural criteria. However, the structural and probabilistic measures complement each other; both are essential in a diagnostic system. We intend to develop probabilistic measures for evaluating coherent hypotheses based on techniques described in [Duda76] and [Pednault81].

The diagnostic problem solver in ABEL has a simple tree structured plan for controlling its diagnostic information gathering. Although it already provides the rudimentary abilities discussed above, it fails to capture the interactions between different branches of the tree. Additional inadequacies arise for the following two reasons. First, as discussed in chapter 3, the use of available knowledge of anatomy, etiology and disease taxonomy is limited. Second, the program does not ascertain the overall state of the patient's health, e.g., the vital signs, stability etc.36 This assessment is an important component of the physician's evaluation and has considerable influence on his formulation of diagnostic goals and strategies. We believe that a similar assessment of the overall state of a patient's health should be modeled in the PSM explicitly, and used in guiding the diagnostic exploration. In coming years we envision implementing diagnostic reasoners with increasing sophistication based on the models of causal reasoning developed in this thesis and on recent advances in planning paradigms (e.g., [Sacerdoti75, Stefik81]).

A serious criticism of the work presented here could be the small size of the domain and the availability of a well defined methods for the initial formulation of the diagnostic problem. This leads to the questions; do the techniques presented here scale up? What are the problems if they are applied to medical diagnosis in a larger domain similar to that of PIP or INTERNIST-I?

The exact methods used by ABEL in the initial formulation of diagnostic problems are domain dependent. We believe that use of similar techniques to limit the size of initial problem formulation is common among clinicians [Elstein78, Kassirer78]. We believe that it is important to distinguish between the processing strategies used in the initial formulation and those used during later stages of the diagnostic process. Substantial further research is needed in identifying strategies for initial processing in larger contexts.

We believe, however, that the multi-level causal representation of medical and patient- specific knowledge, and the description-building processes are independent of the size of the data-base. The major difficulty in using these methods lies in the enormity of the knowledge-base that will be required to adequately cover problems of the size tackled by PIP or the INTERNIST-I.

In summary, this thesis has developed a new representational scheme, capable of capturing some of the subtlety and richness of knowledge employed by expert physicians, and we have presented a new form of diagnostic problem solver which avoids some of the problems present in the previous programs. We believe that the research presented in this thesis is a small step in the right direction. Designing expert medical programs is a difficult and challenging task; much work clearly remains before successful and acceptable expert medical systems are a reality.

The road to wisdom? — Well, it's plain
and simple to express:

and err
and err again
but less
and less
and less.

Piet Hein, Grooks

... on to References

This section is part of

Patil, Ramesh S. Causal Representation of Patient Illness for Electrolyte and Acid-Base Diagnosis. MIT Lab for Computer Science Technical Report TR-267. October 1981. Also: Ph.D. Thesis, MIT Dept. of Electrical Engineering and Computer Science.

The document was reconstructed for the Web in April 2002 by Peter Szolovits.