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).


1. All the previous programs allow the entire hypokalemia to be accounted for by diarrhea. In particular, Internist-I after allowing the hypokalemia to be accounted for by diarrhea will not allow hypokalemia to lend any support to the hypothesis of vomiting. PIP, on the other hand, will allow the entire hypokalemia to lend support to the hypothesis of vomiting as well as allowing it to be explained by diarrhea.

2. This does not pose serious difficulty in medical domains where the pathophysiology of diseases is not well developed, because in such a domain a physician relies primarily on his phenomenological knowledge. However, in a domain such as electrolyte and acid-base disturbances we are constantly faced with this problem because, on the one hand, the pathophysiology of the disturbances is well developed, and on the other, the pathophysiology of many of the diseases leading to these disturbances is relatively poorly understood.

3. The program may not, as was the case with some previous programs, put these suggestions “on hold” without reasoning about them until it is ready to ask about them, If the program does not think that the suggestion is relevant, it must make that decision explicitly.

4. If a disease (A) and one of its inferiors (B) are evoked simultaneously, then (1) if there are no known findings that can differentiate between B and any of its sibling hypotheses, B is considered to be subsumed by A and deleted from the active set. Otherwise, (2) A is replaced by the sot of its immediate inferior diseases that are evoked by the manifestation. Internist -I and Present Illness Program 24

5. During this phase the program does not attempt to identify diseases responsible for the presence of those states. The diagnosis is attempted separately after the information gathering phase is completed.

6. Locality has been exploited in a large number of diverse problems, such as common-sense reasoning [Minsky73, Kuipers77, deKleer79] and natural language processing [Marcus79, Church80, Martin81]. For example, the constraint of “context freeness” in natural language is a specific instance of locality constraint.

7. The numbers corresponding to the acid-base disturbances computed above are the program's internal assessment of the severity of illness; they are not measurable.

8. The data are expressed in XLMS [Hawkinson80], which is briefly described in Appendix I.

9. Each hierarchy, such as the anatomic taxonomy, provides us with a tree structured partial order. The tree structure for the disease definitions is then derived from these partial orders.

10. In the current implementation of ABEL, this knowledge is used only for grouping different findings and diseases in diagnostic problem solving. However, the knowledge representation described in this chapter is capable or supporting a substantially wider variety of uses.

11. A serum creatinine of 1.2 mg per cent can be interpreted in more than one way. For example we can assume this to be normal for a muscular male patient. gut, for a average built female patient this could be an early indication of loss of as much as 1 /3 of the renal function.

12. The number live does not have any medical or cognitive significance; it was chosen for purely engineering reasons.

13. This is typical in English, where the level of detail of place names, for example, is often obtained from context and not encoded in the name used.

14. The causal knowledge described here is encoded by hand, and represents the program's general medical knowledge. A similar multi-level description built by the program to describe a specific patient illness will be discussed in chapter 4.

15. During the exploration of a diagnostic space, traversing a compiled link is equivalent to traversing the predefined path associated with the compiled link in a single step.

16. This is similar to “triggering” a disease in PIP.

17. The use of PSD tree is similar to the use of a “context tree” in CONNIVER [Sussman72].

18. This is Consistent with our view that “the belief in an explanation is equal to the belief of its weakest link”. This belief computation is similar to that used in Glaucoma/CASNET program [Weiss78] and in fuzzy set theory [Zadeh65, Gaines76].

19. A similar distinction is also made in PIP and Internist, where any disease which is not currently active can be considered to be irrelevant to the Current diagnostic activity.

20. On the other hand, de novo generation of the hypothesis list prevents the program from taking “garden paths”.

21. For example, in order to differentiate between alternative hypotheses contained in a DC the program may create a sot of disjunctive DCs, one for each alternative.

22. It can be done if we compute the smallest number of etiologies that cover all unaccounted findings (using the DC) in each PSM before scoring them. This measure, however, has not been implemented as it is computationally prohibitive in the current implementation of the program.

23. Note that a finding that is fully accounted for in the PSM can still be consistent with the new hypothesis if the addition of the hypothesis does not cause the finding to be over-compensated, resulting in an unaccounted component.

24. In certain situations though, the general syntactic mechanism may be overruled by more important considerations. For example, if one of the alternatives has life-threatening consequences, we may first want to get a definitive ruling on it rather than differentiating among all the possibilities.

25. An example of this is the review of systems, a detailed exploration of every part of the body in search of abnormalities.

26. We are using an unrealistically simple example for the purpose of illustration. For this example we have assumed that the patient has received fair quantity of IV fluid. Furthermore, we assume that the electrolyte Concentrations in urine am not available; the differentiation is trivial it the urine electrolytes are available.

27. We have just begun to exploit all the capabilities afforded by this mechanism. Although the current program does not make sophisticated use of these capabilities, we believe that extending the program to do so is possible given the current understanding of the process.

28. Triple disturbances, although possible, are rare and should be considered only when sufficient evidence demands consideration of triple disturbance, generally after one of the components has been confirmed and the acid-base profile after compensating for the known disturbance still requires at least two further disturbances for proper accounting. Quadruple disturbances are almost never considered in clinical practice.

29. Note here that since we are at the pathophysiological level, each link being projected is primitive. Thus, projecting back a node at this level is equivalent to instantiating the cause and the link connecting the cause and the effect node.

30. A node is aggregable if in the medical knowledge-base it is the focus of the elaboration structure of some node at the next higher level which can be instantiated within the PSM. Otherwise, the node is not aggregable.

31. The search terminates when the program finds the first aggregable node on each path.

32. These three hypotheses could be differentiated very easily on the basis of history and clinical evidence. For the simplicity of the example, we assume that this information is not available to the program.

33. A simple criterion for confirming a disease similar to that in PIP or MYCIN can easily be added to the program. However, we have chosen not to do so because of two reasons: first, because the choice of threshold for confirming a disease is arbitrary and therefore, very difficult to explain, and second, in the electrolyte and acid-base program we envision this to be the task of the global decision-making module.

34. The loss of H+ from the extracellular fluid can be viewed as gain in HCO3, because as the H+ is removed from the carbonic acid-bicarbonate buffer an equivalent amount of HCO3 is released into the fluid.

35. The explanation technique developed by Swartout explores the use of automatic programming for encoding a performance program's domain knowledge and principles which are then used to explain the behavior of the performance program. ABEL, however, maintains an explicit account of its knowledge. Therefore, the use of automatic programming is not necessary to explain ABEL's reasoning or understanding.

36. We have often noted clinicians describing a patient in terms such as “this is an otherwise healthy patient with chronic urinary tract infection” or “this is a very sick patient with acute bowel inflammation”.

37. In the medical knowledge base a causal link is interpreted as indicating a possible causal relation.


Appendix I: XLMS

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.