Interdisciplinary research involving artificial intelligence, expert systems, medical informatics, and decision analysis is presented. The book consists of six chapters and five  appendices. 

Chapter 1 introduces the basic concept of normative expert systems that use a decision-theoretic model as the framework for knowledge representation and inference. The advantages of such an approach are highlighted. The breaking point in building normative expert systems has been a graphical knowledge representation called the influence diagram. The author’s contribution to the field is two extensions to the influence diagram representation, similarity networks and partitions. A similarity network is a tool for constructing an influence diagram, and a partition is a tool for assessing the probabilities associated with an influence diagram.

Chapter 2 presents the use of similarity networks and partition representations to construct a knowledge map for a small expert system in the domain of laryngology. This chapter serves as an introduction and explanation of basic ideas such as similarity networks consisting of similarity graphs and a collection of local knowledge maps.

Heckermann describes conditions related to similarity networks and knowledge maps, such as soundness, consistency, and validity. The constraints necessary for the soundness result to hold for any similarity network are listed. He uses assertions of subset independence and hypothesis-specific independence to simplify the assessment of probabilities associated with a knowledge map. In particular, a partition representation is a tool for the assessment of these probabilities.

In chapter 3, a formal theory for similarity networks is developed. In chapter 4, the practical application of similarity networks and partitions is presented using the Pathfinder project expert system, which addresses the domain of lymph node disease. A brief history of the Pathfinder project is given, involving four versions of the system. The most interesting are version III, based on a simple Bayesian model, in which all features are assumed to be independent, and version IV, which tackles the problem of representing dependencies among features. New representations were used in the latest version. Finally, the inference algorithm of Pathfinder is presented.

Chapter 5 compares the diagnostic accuracy of version III and version IV of Pathfinder on 53 cases of lymph node disease. This comparison is done in three steps, using expert ratings, a case-by-case analysis, and a decision-theoretic approach. Chapter 6 gives conclusions and discusses possibilities for future work. Appendix A contains a discussion of basic concepts from decision theory and a tutorial on knowledge maps and influence diagrams. Appendix B contains proofs of theorems.

The basic purpose of this doctoral dissertation is to present research in knowledge acquisition and knowledge representation in an expert system. The results of the research are the new representations, which greatly facilitate the building of an expert system in a complex knowledge domain. Also, the computer implementation, SimNet, a tool for building normative expert systems, is demonstrated. The book can also be treated as a tutorial on the knowledge representations developed by the author, specifically similarity networks and partitions. The book fulfills its basic purpose, and its length is suitable.

The best and most important features of the book are that it focuses attention on building normative expert systems, that is, systems that satisfy the set of constraints imposed by decision theory, thus greatly improving the quality of knowledge; develops extensions of graphical knowledge representation in the form of similarity networks and partitions, which make a large amount of knowledge manageable; verifies these concepts in practice by building the Pathfinder expert system; and gives the historical background of expert systems development and shows the path that led to the idea of the normative expert system.

An important aspect of the book is that it reviews other approaches to clinical decision making under uncertainty, such as Dempster-Shafer theory of belief functions, fuzzy set  theory,  and certainty factor models. The author chooses the decision theoretic framework for his work, and this choice is well founded. Its other important aspect is that new representations have been used to build normative expert systems in domains distant from medicine, proving their domain independence. One criticism may be that the book lacks a formal treatment of similarity, giving only its operational definition.

My overall assessment of the book is positive. The material is presented clearly, and the areas of interest to different types of readers are indicated. The spectrum of potential readers is broad, including researchers in artificial intelligence, knowledge engineers working in medical and other domains, physicians, and medical students.

The list of references is rich and includes approximately 150 books and publications. It could contain more publications on graph theory, as the book relies heavily on that domain. The figures are valuable, and the figure legends allow readers to trace the content of the figures without referring to the text.

I recommend the book strongly. The doctoral dissertation it contains was given an award in 1990 by the Association for Computing Machinery.