Causal Representation of Patient Illness
for
Electrolyte and Acid-Base Diagnosis

This research was supported (in part) by the National Institutes of Health Grant No. 1 P01 LM 03374-Ox from the National Library of Medicine.

Submitted to the Department of Electrical Engineering and Computer Science on October 24,1981 in partial fulfillment of the requirements for the Degree of Doctor of Philosophy

Abstract

Much of the medical knowledge in the first generation Al in Medicine programs is phenomenological; that is, it describes the associations among phenomena without knowledge of the underlying causal mechanisms. Although these AIM programs provide a good first approximation to the way clinicians reason, they fail to reproduce clinicians' reasoning based on a deeper understanding of the phenomena. More specifically, they do not deal with the knowledge of disease at different levels of detail, nor do they utilize causal relations to organize and explain the clinical facts and disease hypotheses. They also cannot deal with illnesses resulting from multiple diseases, especially when one disease alters the presentation of the others. Finally, they are unable to capture the notions of adequacy and parsimony that play such a large role in diagnosis. To explore these issues and rectify these deficiencies, we have undertaken the task of providing expert consultation for electrolyte and acid-base disturbances.

This thesis reports the implementation of ABEL, the diagnostic component of the consultation program. In it, we explore the problems of modeling the causal understanding of a patient's illness. We develop techniques for dealing with illness resulting from multiple interacting diseases. We describe a multi-level representation of causal knowledge, and explore issues of the aggregation of available case specific knowledge into concise summaries of the patient's illness. We discuss structural criteria for evaluating parsimony, coherence and adequacy of diagnostic explanations. We also explore some of the issues involved in information gathering and propose expectation-driven diagnostic planning as a means of improving it. Finally, we discuss the issues of explanation and justification of the program's understanding and argue that these facilities are crucial for acceptability of a consultation program.

Thesis supervisor: Peter Szolovits
Associate Professor of Electrical Engineering and Computer Science

Keywords: Causal Representation, Medical Diagnosis, Structure Building Operations

CONTENTS

1. Introduction	8
1.1 Scope of Project	12
1.2 Choice of Domain	14
1.3 Brief review of Electrolyte and Acid-Base Disorders	15
1.4 Desiderata	19
1.4.1 Making a Correct Diagnosis	19
1.4.2 Continued Management of the Patient	19
1.4.3 Diagnostic Style	19
1.4.4 Mode of Interaction	20
1.4.5 Handling Discrepant Information	20
1.4.6 Explanation	21
1.5 Survey of AIM programs	22
1.5.1 Internist-I and Present Illness Program	23
1.5.2 CASNET/Glaucoma	25
1.5.3 Mycin	25
1.6 Outline of the Thesis	27
2. Examples	30
2.1 Example 1: Salmonellosis	30
2.2 Example 2: Vomiting and Salmonellosis	40
3. Representation of Medical Knowledge	43
3.1 Anatomical Knowledge	44
3.1.1 Anatomical Taxonomy	44
3.1.2 Material Flow Pathways	44
3.1.3 Anatomical Spaces	46
3.1.4 Miscellaneous Gross Anatomical Relations	48
3.2 Etiological Knowledge	49
3.3 Physiological Knowledge	50
3.4 Disease Knowledge	51
3.5Causal Link	53
3.6 Multi-Level Causal Description	54
4. Structure Building Operations	61
4.1 Structure of a PSM	61
4.2 Initial Formulation	62
4.3 Some Definitions	63
4.4 Aggregation	65
4.4.1 Focal Aggregation	65
4.4.2 Causal Aggregation	66
4.5 Elaboration	69
4.5.1 Focal Elaboration	69
4.5.2 Causal Elaboration	70
4.6 Projection	72
4.7 Component Summation and Decomposition	72
5. Diagnostic Problem Formulation and Information Gathering	80
5.1 Global Diagnostic Cycle	82
5.2 Diagnostic Closure of a Hypothesis	83
5.3 Scoring the PSM	85
5.4 Scoring a Disease Hypothesis	85
5.5 Information Gathering Strategy	88
6. Examples Revisited	95
7. Conclusion	118
7.1 Summary	118
7.2 Limitations of ABEL and Future Directions	119
8. References	122
Notes
Appendix I. System Building Tool: XLMS	130
I.1 XLMS Concepts	130
I.2 The XLMS Interpreter	133
Appendix II. Explanation	134
II.1 Phrase Generator	135
II.2 Higher level explanations	135
II.3 Organizing causal Explanation	136

FIGURES

Fig. 1. A schematic for the overall system	13
Fig. 2. Carbonic acid - bicarbonate buffer equation	17
Fig. 3. Nomogram of acid-base disturbances	18
Fig. 4. Graphic depiction of the two Acid-Base hypotheses	31
Fig. 5. Comparison of hypotheses 1 & 2 at clinical level	33
Fig. 6. Comparison of hypotheses 1 & 2 at intermediate level	34
Fig. 7. The part-of hierarchy	45
Fig. 8. Material flow relations	46
Fig. 9. The containment relation	47
Fig. 10. Gross anatomical relations	48
Fig. 11. Etiological hierarchy	49
Fig. 12. Schematic description of a causal link	53
Fig. 13. Schematic description of the node structure	55
Fig. 14. Comparison of lower Gi fluid and of plasma	56
Fig. 15. The loss of electrolytes in lower Gi fluid	56
Fig. 16. Consequences of lower Gi loss described at next higher level	57
Fig. 17. Lower Gi loss expressed at an intermediate level	57
Fig. 18. Salmonellosis and its consequences expressed at the clinical level	58
Fig. 19. Layered description of link: salmonellosis causes dehydration	58
Fig. 20. Compiled link	59
Fig. 21. Node types	64
Fig. 22. Causal aggregation: fully Unaccounted node	66
Fig. 23. Causal aggregation: fully accounted node	67
Fig. 24. Causal aggregation: partially accounted node	68
Fig. 25. An example of the elaboration process	71
Fig. 26. An example of component summation/decomposition	73
Fig. 27. Feedback loop represented using component summation	74
Fig. 28. Component summation/decomposition: Case 3	77
Fig. 28. (continued)	78
Fig. 29. An example of diagnostic closure	84
Fig. 30. An example of explained, unexplained and unaccounted findings	86
Fig. 31. Initial diagnostic closure for salmonellosis and acute renal failure	89
Fig. 32. Diagnostic closure separated for each possibility	90
Fig. 33. Diagnostic closures for each possibility projected forward	91
Fig. 34. The goal tree	92
Fig. 35. Serum electrolytes and the bar diagram	95
Fig. 36. Graphical description of acid-base disturbances	96
Fig. 37. Hypothesis 1	98
Fig. 38. Hypothesis 2	99
Fig. 39. Aggregation of low-serum-K-1	100
Fig. 40. Aggregation of low-pH-1	100
Fig. 41. PSM for hypothesis 1	101
Fig. 42. PSM for hypothesis 2	102
Fig. 43. Diagnostic closure 1	103
Fig. 44. Diagnostic closure 2	104
Fig. 45. One complete cycle of diagnostic inquiry	105
Fig. 46. Diagnostic closure 3	106
Fig. 47. Diagnostic closure 4	107
Fig. 48. Diagnostic closure 5	107
Fig. 49. After all findings have been exhausted	108
Fig. 50. Hypothesis 1 with salmonellosis	109
Fig. 51. Hypothesis 2 with salmonellosis	110
Fig. 52. English description of the two hypotheses	112
Fig. 53. Initial PSM	114
Fig. 54. Revised PSM after vomiting is entered	115
Fig. 55. Final PSM after salmonellosis is introduced	116
Fig. 56. English text of the final explanation	117
Fig. 57. The XLMS hierarchy	131
Fig. 58. Feedback relation betweeen acidemia and hypocapnia	137

Acknowledgments

I would like to express my thanks to all of the people who made this thesis possible:

Peter Szolovits, my thesis supervisor, for providing constant attention, guidance and for helping me formalize ideas when I could not see through the confusion;

William B. Schwartz for suggesting this thesis topic, teaching me about acid-base and electrolyte disturbances, and for articulating and refining many of the ideas presented in this thesis;

Randall Davis, for his helpful comments and suggestions on drafts of this document;

William Martin for introducing me to the area of knowledge based application systems and providing the intellectual environment in which these ideas germinated;

Lowell Hawkinson for developing XLMS and providing help in using it;

Bill Long for his encouragement and for being a very insightful sounding-board for ideas;

Glenn Burke for proofreading this document and for providing much needed system support;

Ken Church for many spirited arguments and discussions on this and other topics;

Bill Swartout for timely development of explanation methodology and for sharing his experience in using XLMS;

Stephen Pauker, Brian Smith, Ben Kuipers, Byron Davies, Howard Sherman, Harold Goldberger, Brij Masand, Gretchen Brown and other past and present members of Clinical Decision-Making group and Knowledge-Base System groups for providing the fertile environment and comradery needed to carry me through the thesis;

and finally, my wife Aruna and my family for bearing with me and providing constant encouragement without which this would not have been possible.

... on to Chapter 1