A Survey on Case-based Reasoning in Medicine

Case-based reasoning (CBR) based on the memorycentered cognitive model is a strategy that focuses on how people learn a new skill or how they generate hypothesis on new situations based on their past experiences. Among various Artificial Intelligence tracks, CBR, due to its intrinsic similarity with the human reasoning process has been very promising in the utilization of intelligent systems in various domains, in particular in the domain of medicine. In this paper, we extensively survey the literature on CBR systems that are used in the medical domain over the past few decades. We also discuss the difficulties of implementing CBR in medicine and outline opportunities for future work. Keywords—case-based reasoning; medicine; artificial intelligence; soft computing


INTRODUCTION
Case-Based Reasoning (CBR) is an area of machine learning research based on the memory-centered cognitive model [1].CBR arose out of the research in cognitive science.It is defined as a model of reasoning that integrates problem solving, understanding and learning, and incorporates all of them with memory processes.It involves adapting earlier solutions to meet new demands, using old cases to explain or justify new solutions, and reasoning from past events to interpret a new situation.In CBR terminology, a case usually denotes a problem situation [2].CBR can be considered as a form of similarity-based or analogical reasoning since the basic principle that is implicitly assumed to be applied in problem solving methodology is that similar problems have similar solutions [3].
CBR as a problem solving paradigm, is essentially different from other major Artificial Intelligence (AI) approaches in many aspects.Unlike other approaches which rely solely on the general knowledge of a problem domain, or which associate along inferred relationships between problem descriptors and conclusions, CBR utilizes the specific knowledge of previously experienced problem situations [2].CBR can be applied as "reasoning by experience in AI" as compared to rule-based reasoning which is applied as "reasoning by logic in AI" [4].The intuitive appeal of CBR comes due to its similarity to human problem solving behavior.Just as people draw on past experiences while solving a new problem, which often does not require in-depth analysis of the problem domain, CBR can be based on shallow knowledge and does not require significant effort in knowledge engineering as required by other AI fields like rule-based reasoning [5].
Medical reasoning on the other hand, involves processes "that can be systematically analyzed, as well as those characterized as intangible" [6].In medicine, the experts not only use rules to diagnose a problem, but they also use a mixture of textbook knowledge and experience.The experience consists of cases, typical and exceptional ones, and the physicians take them into account for reasoning.So, caseoriented methods should be very efficient in the domain of medical diagnosis, mainly because reasoning with cases corresponds with the typical decision making process of physicians.Also, incorporating new cases means automatically updating parts of the changeable knowledge [7].Despite these, CBR has not become as successful in the medical domain, as it is in other fields for building intelligent systems [8].
The present paper surveys the available literature on systems developed using CBR for solving various problems in medicine.We begin in Section 2 by describing the basic notions of CBR and its models, with a brief description of the phases in CBR life cycle.Section 3 gives a brief description of medical reasoning.Section 4 surveys various CBR based systems developed over past few decades in the domain of medicine.In Section 5, we point out certain issues of using CBR in the field of medicine.Section 6 concludes the paper with a discussion on future directions of research.

II. INSIDE CASE-BASED REASONING
CBR is an analogical reasoning method, which means that it reasons from old cases or experiences to solve problems or interpret anomalous situations [9].But the major difference between CBR and analogy is that analogy reasons across domains, whereas CBR reasons inside one domain [10].In CBR, the reasoning is based on remembering past experiences, as explained by Althoff et al. [11] -"To solve a problem, remember a similar problem you have solved in the past and adapt the old solution to solve the new problem."CBR can be interpreted in many ways [12] by different groups of people.For example, for cognitive scientists, it is a plausible high-level model for cognitive processing; for artificial intelligence researchers, it is a computational paradigm for solving problems; and for expert system practitioners, it is a design model.
CBR arose out of the research in cognitive science.The earliest contributions in this area were from Roger Schank and his colleagues at Yale University [2].During the period 1977-1993, CBR research was regarded as a plausible high-level model for cognitive processing.Three CBR workshops were organized in 1988, 1989, and 1991 by the U.S. Defense Advanced Research Projects Agency (DARPA), which officially marked the birth of the discipline of CBR.In 1993, the first European workshop on CBR (EWCBR-93) was held www.ijacsa.thesai.org in Kaiserslautern, Germany; and the first International Conference on CBR (ICCBR-95) was held in Sesimbra, Portugal.Many international workshops and conferences on CBR have been held in different parts of the world since then.Medical applications have been a part of the CBR community from the very beginning and are included in almost every international conference on CBR [13].

A. CBR Models
To understand the working of CBR, various models have been proposed in the literature.These include Hunt"s model, Allen"s model, Kolodner and Leake"s model [14], and R 4 model, developed by Aamodt & Plaza [2].Of these, the most widely used model and at the highest level of generality is the R 4 model [15].The process involved this model can be represented by a schematic cycle comprising of the four R"s, as illustrated in Figure 1.

B. CBR Life Cycle
The problem solving life cycle of CBR essentially consists of retrieval, adaptation, and maintenance.Each of these has its own importance in the successful working of the CBR system.

1) Retrieval
Retrieval is often considered the most important phase of CBR since it lays the foundation for overall working of the CBR system [16].Retrieval includes the process of finding those cases within a case base, which are most similar to the current case.The most commonly investigated retrieval techniques include nearest neighbor retrieval, inductive approaches, knowledge guided approaches, and validated retrieval [5], [17].Some hybrid algorithms have also been proposed e.g.Discretised Highest Similarity with Pattern Solution Re-use algorithm [18].

2) Adaptation
The next two phases of the CBR cycle, viz.reuse and revise are often difficult to distinguish in many practical applications, as a result of which many researchers replace and combine them into a single stage called adaptation [5].In the early 90"s the CBR community focused on retrieval only.Investigations of the various aspects of adaptation started after that [19].Most of the advances also have been achieved at the retrieval and retain phase of CBR [20].In the reuse phase, advances have been obtained depending on the system purpose viz.diagnosis, classification, tutoring and planning (such as therapy support).Regarding diagnosis and classification, most of the systems rely on adaptation methods that consist of copying the solution of the most similar case or a combination of them, i.e. reusing the solution [21].

3) Maintenance
After reusing and revising the retrieved case, the next step in CBR cycle is to retain the case (s).There are many approaches to achieve this.Many systems store only the solution of the previous problem, whereas some systems store the solving process [16].In many cases, this process of retaining leads to an uncontrolled growth in the case base, which in turn leads to a poorer performance of the system in terms of speed [22].So, the need of maintaining a case base arises.

III. MEDICAL REASONING
Medical reasoning is divided into diagnostic reasoning, planning, and patient management [23].This reasoning is carried out in terms of physiological states, complaints, symptoms and so forth [24].Diagnostic reasoning includes cognitive activities like gathering information, recognition of patterns, solving problems and decision making [25].Diagnostic investigations are quite complex and error prone [26].Table 1 outlines the diagnostic process.DIAGNOSTIC CYCLE [27] This diagnosis process may become easier and more reliable if equipped with an expert system that provides past diagnosis of cases, thereby helping the physician to arrive at a solution based on the past experiences [28].

IV. CBR SYSTEMS IN MEDICINE
CBR used in medical reasoning literature is termed as "instance-based recognition" [29].Unlike other knowledge domains, cases have to be professionally documented in medical domain [30].The very fact that the methodology of CBR systems closely resembles the thought process of a physician suggests a successful use of CBR in medicine [31].Koton pointed out while introducing CASEY -"A physician"s problem-solving performance improves with experience.The performance of most medical expert systems does not" [32].The experts in the medical domain do not use rules for diagnosis.What they use is the knowledge they obtain from books, as well as experiences just the way in which CBR works [7].
Step Decision 1 Select a diagnostic test (or question) 2 Carry out the selected test and observe its outcome 3 Either (i) select a further diagnostic test and so return to step 1; or (ii) make a diagnosis in the light of the outcomes so far obtained www.ijacsa.thesai.org The main advantage of CBR systems in medicine is the automatic formation of a facility adapted knowledge base [33], which is a very important aspect in medical decision making.Also, the continuously changing nature of medical knowledge base, presence of more than one solution, and complexity in modeling also make CBR applicable in medical domain [34].As a result, CBR has been used for building intelligent computer-aided decision support systems in the medical domain in the past few decades [35].
CBR decision support systems can be classified [20] as planning, classification, tutoring, and diagnostic systems based on their purpose oriented properties.Table 2 lists in chronological order, some of the CBR systems developed in the field of medical reasoning over the years.Also, it classifies these systems according to their objectives and attempts to find out the extent to which adaptation phase of CBR is used in these systems.From our study, it was observed that CBR in the medical domain has a wide range of application.Most of the systems are developed specifically to deal with a particular disease.Secondly, most of the systems act as prototypes, and not as final products, as mentioned by Blanco [110].These systems require a human expert to interpret the final result.Another visible trend was the successful hybridization of CBR with soft computing methods.32 out of 76 systems studied by us have used some or the other soft computing techniques in addition to CBR.Moreover, among the 76 systems, 51 systems completely avoid automatic adaptation and mainly work as retrieval only systems.The other systems do have the adaptation phase in them, but often the reasoning mechanism in those is coupled with rule-based reasoning, or various soft computing methods.www.ijacsa.thesai.orgThough the above discussion reflects the successful use of CBR in medicine, there are some limitations which restrict the use of CBR in medicine.In a medical case, the number of features is often extremely large, thereby making the generalization and adaptation quite difficult [20].At the same time, reliability cannot be guaranteed in medical CBR systems [111].The limited number of reference cases aids to the problem of implementing a medical CBR system [35].But the most important concern in the successful implementation of medical CBR systems is the adaptation problem.As our study suggests, so far, the number of systems in the medical domain that apply the complete CBR method is very less.Most of the systems use no adaptation at all, and the task of adaptation is left to the human expert.
d"Aquin et al. [59] remark that adaptation in medicine is quite a complex procedure, as it needs to deal with the lack of relevant information about a patient, the applicability and consequences of the decision, the closeness to the decision thresholds and the necessity to consider patients according to different viewpoints.Schmidt et al. [7] also point out that giving autonomy to the adaptation step of CBR has been a difficult step in Medicine.Due to these challenges, most of the advances made in medical CBR systems focus on the retrieval phase.The adaptation phase is limited to planning tasks [21].No general models have been developed for adaptation as it largely depends on the domain and application characteristics.
Our study reveals that medical CBR systems deal with the adaptation problem in two ways.Most of the systems avoid the adaptation problem by applying only retrieval phase of CBR cycle [19] while some others attempt to solve it.One of the earliest medical expert systems, CASEY [32] makes an attempt to solve the adaptation task.In this, the creation of a complete rule base for adaptation is time consuming, as a result of which a few general operators are used for adaptation.And when no similar case can be found or if adaptation fails, CASEY uses a rule-based domain theory.But since knowledge acquisition is the bottleneck for the development of rule-based medical expert systems, the development of complete adaptation rule bases have never become a successful technique to solve the adaptation problem in medical CBR systems [7].The application of constraints leads to a better solution, as in the GS.52 project [37] but only for specific situations.KASIMIR [59] uses similarity paths and reformulation to support the adaptation, but adaptation knowledge in the form of rules is still required.Some of the more recent systems perform adaptation successfully, with the help of soft computing techniques, e.g.eXiT*CBR.v2[98] revises and reuses the cases using genetic algorithms; EquiVox developed by Henriet et al. [101] performs adaptation using artificial neural networks.So, the inclusion of soft computing techniques suggests improved automatic adaptation in medical CBR systems.

VI. CONCLUSIONS AND FUTURE SCOPE
A fundamental part of the CBR system is learning by remembering cases.CBR systems, cognitively similar to human beings, take into account previous experiences for solving new problems, consider both subjective and objective knowledge unlike other expert systems, and can incrementally acquire knowledge automatically, but still, these are not as successful in medicine as in other domains.The main reason for this is the adaptation problem.The retrieval and maintenance phases have gained a lot of attention of the researchers, while the phase is still in its infancy.The adaptation phase involves multifarious problems which include dealing with the closeness to the decision threshold used to determine similar cases, among other issues.The majority of the medical CBR systems avoid the adaptation problem, and act as retrieval only systems and leave case adaptation and case update to be performed by human experts.A solution to adaptation problem is the integration of CBR with other methodologies.The synergism of these methodologies leads to the development of new sophisticated and hybridized systems.
It was observed in our survey that a majority of successful medical CBR systems are built around a combination of CBR and other artificial intelligence methods.From the very beginning, hybrid systems came into existence for medical CBR systems; Koton"s CASEY [32] being an example which hybridizes CBR and RBR.Soft computing techniques viz.fuzzy logic, artificial neural networks, in particular backpropagation neural networks and Bayesian models, and evolutionary strategies have proved to be very efficient in enhancing the capabilities of CBR systems.With the use of these techniques, adaptation knowledge can be determined automatically from the cases, which leads to more robustness of this knowledge [5].Schmidt, Vorobieva, & Gierl [8] have mentioned that the application of adaptation rules or operators, though general seems to be the only technique which can solve medical adaptation problems.We suggest the use of fuzzy decision trees for this; wherein fuzzy decision rules can be generated, and rough set techniques can be used to simplify these rules.
In the domain of medicine, where clear domain knowledge is often not available, automatic adaptation is difficult to develop.So, hybrid combinations of soft computing techniques may be explored and implemented in greater details in the adaptation phase of CBR to move forward the success story of CBR in the otherwise difficult domain of medicine.

Fig. 1 .
Fig. 1.The R4 Cycle [2]  Retrieve the most similar case (s)  Reuse the information and knowledge from retrieved case (s) to solve the problem  Revise the proposed solution  Retain the parts of this solution likely to be useful in future.

Fig. 2 .
Fig. 2. Percentage of systems in terms of adaptation performed V. PROBLEMS IN MEDICAL CBR SYSTEMS