Hierarchical representation of differential diagnosis lists for clinical decision support systems Takashi Okumura National Institute of Public Health, Japan 2-3-6 Minami, Wako city Saitama Pref., Japan taka@niph.go.jp Masaki Tagawa Kyushu Institute of Technology 680-4 Kawazu, Iizuka city Fukuoka Pref., Japan masaki@club.kyutech.ac.jp ABSTRACT Clinical decision support systems (CDSSs) can generate dif- ferential diagnosis lists that may contain hundreds of dis- eases. These lists grow in size as coverage expands to rare diseases, but large lists can easily become a burden on user cognition. To address this issue, we first outline the repre- sentations of differential diagnosis lists on current CDSSs, and then propose a novel approach that represents these differential diagnosis lists hierarchically, coupled with an al- gorithm for optimal initialization. Preliminary evaluation suggested that our proposed approach outperforms exist- ing approaches with respect to search costs, particularly for large lists. This hierarchical representation should alleviate the cognitive load on user physicians and provide an efficient means to search through very large lists. Categories and Subject Descriptors H.5.2 [Information Systems]: Information interfaces and presentation General Terms Clinical Decision Support Systems, Expert systems Keywords Differential diagnosis, Hierarchical representation 1. INTRODUCTION A differential diagnosis list is a list of possible diagnosis that is tentatively built for planning further examinations aimed toward an accurate diagnosis. Physicians are educated to consider at least three possible diagnoses in Japan: com- mon, curable, and critical diseases, and these three can be a minimal set for differential diagnosis. For most situations, however, there are several more possibilities to be consid- ered. Nevertheless, a list should not exceed 20 or 30 items, because this exceeds the capacity of human cognition. In these cases, physicians set up a diagnostic hypothesis to fo- cus on a more limited number of possibilities. In contrast, clinical decision support systems (CDSS) can generate a differential diagnosis list with a far greater vol- ume. This list is built on a set of given findings, and pro- vides a diagnostic probability for each item. The resulting list is an unstructured flat list, sorted by the probabilities, although such a list becomes too monotonous to browse as the number of listed items grows. This situation occurs dur- ing the diagnosis of rare diseases, if physicians have no choice but to use general search engines by providing key findings of a case as their search terms [10]. In these cases, physi- cians are forced to laboriously search through a large list of either disease names, or web pages. To develop a CDSS that covers various rare diseases, im- proving the representation would be indispensable. Accord- ingly, this article addresses this problem, through a hier- archical representation of a differential diagnosis list. This structured representation should alleviate the cognitive load placed on user physicians and provide an efficient means to formulate and test their clinical hypotheses. This should also contribute to the usability of related health-care appli- cations. The sections in this article are organized as follows. Sec- tion 2 provides a survey of existing representations on rel- evant CDSSs. Section 3 provides the hierarchical represen- tations of differential diagnosis lists, and an algorithm for efficient initialization. Section 4 demonstrates the results of our preliminary evaluation. Section 5 discusses the proposed representation, and Section 6 concludes this paper. 2. BACKGROUND AND RELATED WORK Although there are a various types of CDSSs [2], which are not limited to those that output differential diagnosis lists, they form the mainstream of research and production sys- tems. Many systems in this category generate flat lists of possible diagnoses that are sorted by their possibilities (Fig- ure 1-a), using given sets of clinical findings as their inputs [11, 12]. Such a simple listing is a good fit, particularly for people without clinical knowledge [12], as the list only in- cludes a limited set of common diseases and this simplicity is well suited for such users. Even experts may prefer a flat listing if the context is limited, e.g., the diagnosis of infec- tious diseases only [11]. However, for hard-to-diagnose cases, the situation changes, because a CDSS may generate hun- dreds of possible diagnoses. Such a list can easily overwhelm human cognition if the candidate diseases are presented in a simple list that requires a serial search. 3HUPLVVLRQ WR PDNH GLJLWDO RU KDUG FRSLHV RI DOO RU SDUW RI WKLV ZRUN IRU SHUVRQDO RU FODVVURRP XVH LV JUDQWHG ZLWKRXW IHH SURYLGHG WKDW FRSLHV DUH QRW PDGH RU GLVWULEXWHG IRU SURILW RU FRPPHUFLDO DGYDQWDJH DQG WKDW FRSLHV EHDU WKLV QRWLFH DQG WKH IXOO FLWDWLRQ RQ WKH ILUVW SDJH 7R FRS\ RWKHUZLVH WR UHSXEOLVK WR SRVW RQ VHUYHUV RU WR UHGLVWULEXWH WR OLVWV UHTXLUHV SULRU VSHFLILF SHUPLVVLRQ DQGRU D IHH ,'3+$ 0D\ &RS\ULJKW k ,&67 '2, LFVWSHUYDVLYHKHDOWK