Modeling Paradigms for Medical Diagnostic Decision Support: A Survey

and Future Directions∗

Kavishwar B. Wagholikar Vijayraghavan Sundararajan Ashok W. Deshpande

Email:waghsk@gmail.com

Jan 2011

Abstract

Use of computer based decision tools to aid clinical decision making, has been a primary goal of research inbiomedical informatics. Research in the last five decades has led to the development of Medical Decision Support (MDS)applications using a variety of modeling techniques, for a diverse range of medical decision problems. Thispaper surveys literature on modeling techniques for diagnostic decision support, with a focus on decisionaccuracy. Trends and shortcomings of research in this area are discussed and future directions are provided.

The authors suggest that – i) Improvement in the accuracy of MDS application may be possible bymodeling of vague and temporal data, research on inference algorithms, integration of patient informationfrom diverse sources and improvement in gene profiling algorithms; ii) MDS research would be facilitatedby public release of de-identified medical datasets, and development of opensource data-mining tool kits; iii)Comparative evaluations of different modeling techniques are required to understand characteristics of thetechniques, which can guide developers in choice of technique for a particular medical decision problem; iv)Evaluations of MDS applications in clinical setting are necessary to foster physicians’ utilization of thesedecision aids.

1 Introduction

MDS applications aim to assist physicians in clinical decision making [1]. Physicians are prone to making decisionerrors, because of high complexity of medical problems and due to cognitive limitations. Decision making inmedicine is complex because, a vast amount of knowledge is required even to solve seemingly simple problems [2].A physician is required to remember and apply knowledge of a large array of entities like disease presentations,diagnostic parameters, drug combinations and guidelines [3]. However the physician’s cognitive abilities arerestricted due to factors like multi-tasking, limited reasoning and memory capacity [4, 5]. Consequently, it isimpossible for an unaided physician to make right decisions every-time [6]. Ironically, the increasing rate ofinformation generation by medical advances has aggravated the physician’s task [7, 8]. Humans have a “limitedchannel capacity” [4], due to which they are unable to process large amount of data. These limitations resultin an estimated 30% morbidity and mortality [9, 10], which is avoidable.

The potential of computer based tools for medical decision making was realized half a century ago [11],and several algorithms have been developed to construct MDS applications for a variety of medical speciali-ties. Figure 1 shows the logical components of the diagnostic process, as outlined by Ledley and Lusted [12].Accordingly, MDS applications typically have two components – medical knowledge-base and an inference en-gine. Medical knowledge is composed of symptom-disease, inter-symptom and inter-disease relationships. Theinference engine essentially interprets patient information using medical knowledge to compute decisions.

∗The final publication is available at www.springerlink.com. dx.doi.org/10.1007/s10916-011-9780-4 . Kindly cite this article

as Modeling Paradigms for Medical Diagnostic Decision Support: A Survey and Future Directions. Wagholikar K. B., Sundararajan

V., Deshpande A. W., Journal of medical systems, pp. 1-21, 2011

1

The final publication is available at www.springerlink.com. dx.doi.org/10.1007/s10916-011-9780-4. Kindly cite this article as Modeling Paradigms for Medical Diagnostic Decision Support: A Survey and Future Directions. Wagholikar K. B., Sundararajan V., Deshpande A. W., Journal of medical systems, : Volume 36, Issue 5 (2012), Page 3029-3049

Patient Information

Medical Knowledge

DiagnosisInference

Figure 1: Logical components of the process of medical diagnosis

Despite half a century of research [13, 14], MDS applications have not received wide acceptance and utiliza-tion in health-care, due to the following reasons:

• Decision accuracy: High level of decision accuracy is desirable in MDS and very few applications havematched the diagnostic performance of human experts [15, 16, 17].

• Integration with workflow: The tools do not provide all the different types of information that the cliniciansneed [7, 18]. For instance, a diagnostic tool often does not provide the plan of treatment and the physicianhas to seek this information from other tools, which interrupts his workflow.

• Usability: The interfaces of many of the early systems were not user friendly and training was requiredfor their use. Usability issues limited their use [19, 20].

• Natural language processing: Most patient information is unstructured and in a free textual form. Suchdata is first parsed using Natural Language Processing (NLP) algorithms to identify the parts of speechlike noun, adjective, verb, etc. Ontologies are then used to perform semantic interpretation of the mappedwords. NLP methods are still not adequate for clinical applications [21].

• Explanation: If scientific evidence is provided to explain the basis of decision-suggestions, physicians aremore likely to use the system [22, 23]. Many systems have lacked such explanation modules.

Decision accuracy is a primary factor determining the adoption and utilization of MDS applications in routineclinical practice. This paper presents a survey of modeling techniques for diagnostic decision support, with afocus on decision accuracy. Though MDS modeling techniques have been applied for a range of medical decisionsincluding diagnosis, prognosis and therapy planning, we have restricted the scope of the survey to the diagnosticdecision problem. Diagnosis is the most researched decision, as diagnostic decisions are comparatively easierto validate. Moreover, most medical decision problems are classification problems and results for diagnosticmodeling techniques can be usually extrapolated to other problems.

The aim of this survey is to provide an overview of modeling techniques for diagnostic decision support, alongwith a listing of their applications, and to provide a context and future directions to readers who are new to thefield. The survey is organized by grouping literature on the basis of modeling techniques. Early MDS systemswere largely bayesian, and were followed by fuzzy set theory (FST), rule-based and heuristic system. RecentlyBayesian Network (BN), Artificial Neural Network (ANN) and Support Vector Machine (SVM) models havebeen developed (see figure 8). The next section describes the approach for the literature survey, followed by anoverview of literature of different modeling techniques. The discussion section indicates trends, limitations andfuture directions.

2 Methods

We queried PubMed for Computer assisted Diagnosis (see figure 2) and largely excluded algorithms for image,waveform and -omic data analysis, because each of these areas is a specialized field in itself and applicationsin these domains generally cannot be ported to other problem areas. PubMed search yielded a total of 2880

2

papers. After excluding literature from the specialized areas mentioned above, we analyzed this set to delineatepapers that reported either a new decision model or evaluation of existing models. These were then segregatedinto different groups based on the modeling technique used in the studies. Some reports were left out due topaper length restrictions, yielding the final set of 220 reports cited in this survey.

A limitation of our literature survey is that, our search largely relied on PubMed. Many conference pro-ceedings and journals are not indexed in PubMed and hence, our survey is limited to a subset of publishedliterature.

“diagnosis, computer assisted”[MeSH Major Topic] NOT “Image Processing, Computer-Assisted”[MeSH Terms] NOT

“Biological Markers”[MeSH Terms] NOT “instrumentation”[MeSH Subheading] NOT “Therapeutics”[MeSH Terms] NOT

“Diagnostic Imaging”[MeSH Terms] NOT “Image Interpretation, Computer-Assisted”[MeSH Terms] NOT “Diagnostic

Techniques and Procedures”[MeSH Terms]

Figure 2: PubMed Query

3 Results

This section is organized by grouping literature on the basis of modeling techniques. Early MDS systems werelargely bayesian, and were followed by FST, rule-based and heuristic system. Recently BN, ANN and SVMmodels have been developed (see figure 8).

3.1 Naive Bayes (NB)

Ledley and Lusted’s [12] paper outlining the use of Bayes formula to estimate probability of a diagnosis stim-ulated great interest. This approach is based on the assumption that the findings/symptoms for a particulardisease are not interdependent. For example, the probability that a patient with symptoms of cough and feverhas a diagnosis of puemonia is computed as:

P (pnuemonia)xP (cough|pneumonia)xP (fever|pnuemonia)

P (cough ∩ fever)(1)

where P (cough|pnuemonia) indicates probability of cough given the diagnosis of pneumonia. The assumptionmade for simplification, here is that there is no dependence between fever and cough.

One of the earliest studies with this approach was by Homer Warner at University of Utah. Warner madea probabilistic model to diagnose patients with one of 35 congenital heart diseases [24]. The model was derivedon how frequently 50 different findings occurred in each disease and how common the diseases were in thepopulation of patients referred to his laboratory. Warner’s research led to the development of HELP [25],which was the first hospital information system with decision support modules. Its decision support functionshave been expanded over the years to provide alerts/reminders, data interpretation, patient diagnosis, patientmanagement suggestions and clinical protocols. An offshoot of HELP was a sequential Bayes model of Iliad[26], developed for diagnosis in the domain of internal medicine.

In 1970s, De Dombal developed a system for diagnosis of pain in abdomen at Leeds University, UK. Thiswas one of the first systems to be utilized at clinical sites and is perhaps the most evaluated system, with studiesin UK and over 64 European institutions, together consisting of 37,000 cases [27]. These studies confirmed thatMDS systems could outperform unaided clinicians, but aided clinicians matched and at times exceeded thediagnostic accuracy of the systems.

In 1987, DXPLAIN for internal medicine [28], was developed by Barnett et al. at Massachusetts GeneralHospital. It is currently available as a web-based system and includes 2200 diseases and 5000 symptoms in itsknowledge-base. DXPLAIN produces a ranked list of diagnoses for clinical manifestations of a query patientand also provides justification for why each of these diagnoses might be considered and suggests what furtherclinical information would be useful to collect to further resolve the problem.

Modifications to Bayes rule to overcome the independence assumption, have been applied for diagnosis ofacute abdominal pain [29], sport injuries [30], disease outbreak [31], lung disease [32] and human malformationpatterns [33, 34]. MEDAS system at Chicago medical school used a modified Bayes approach [35].

3

Other early applications of the bayesian approach include diagnosis for hematologic disorders (HEME)[36], and appendicitis [37]. Later applications have been developed for abdominal pain [38, 39, 40, 41, 42],odontogenic lesions [43], oncology [44, 45, 46], liver disorders [47, 48, 49, 50, 51, 52], pancreatitis [53], lung disease[54, 55], dentistry [56, 57], gynecology [58, 59], neurology [60], rheumatology [61, 62], dermatopathology [63, 64],opthalmology [65], hematopathology [66, 67], hypertension [68], heart disease [69], adverse drug reactions [70]and bowel pathology [71, 72, 73].

Many naive bayes (NB) based applications achieved a performance which was comparable to human physi-cians and some were successfully deployed in hospitals [27, 25, 28]. However, the assumptions of independence ofsymptoms and mutual exclusivity of diseases were regarded as over-simplistic, which led researchers to exploreother approaches.

3.2 Decision Analysis

Decision analysis refers to “choice under uncertainty, that is consistent with a set of judgments and preferencesfor consequences”. The preferences for consequences are in terms of utility values, and judgments about uncer-tainties are modeled as probabilities. In clinical medicine, the analysis consists of determining the best action(diagnostic or therapeutic) for any given patient [74]. The physician considers the risks and costs of making thedecisions. For instance, patients complaining of chest pain are routinely screened with Electrocardiogram (ECG)examination for myocardial infarction (heart-attack). Of the many causes for chest pain, myocardial infarctionis one of the less likely ones, but entails high risk of fatality. Hence, the cost of a missed diagnosis of myocardialinfarction is much higher than other causes.

In 1970, Gisnberg [75] and Gorry [76] implemented decision theory model for the diagnosis and treatmentof pleural effusion syndrome and acute renal failure, respectively. Later PIP (Present Illness Program) wasa system built by MIT and Tufts-New England Medical Center, for diagnosing renal disease. It used bothcategorical as well as probabilistic criteria for computing diagnosis [77]. After considerable groundwork toevaluate several approaches, in 1992 Heckerman developed a normative (Pathfinder) system [78] to assist surgicalpathologists with the diagnosis of lymph-node diseases. The system used probability and decision theory toacquire, represent, manipulate, and explain uncertain medical knowledge. Research for Pathfinder led to anunderstanding of the reasons for the failure of some early Artificial Intelligence (AI)-based approaches.

Although physicians use a decision analysis based approach for making clinical decisions, utility considera-tions have been rarely used in MDS applications. A correction in MDS applications’ output by using decisionanalysis, may facilitate clinical utility of the applications.

3.3 Rule-based

Rule-based systems emphasize on use of decision rules for inference, as opposed to scoring functions. One ofthe first and widely known rule-based systems was MYCIN, developed by Shortliffe [79] at Stanford in 1977.Clinical knowledge in this application was represented as a set of if-then rules with certainty factors attachedto consequent diagnoses (see figure 3). A backward chaining reasoning strategy was used for inference.

IF the infection is primary bacteremia

AND the site of the culture is one of the sterile sites

AND the suspected portal of entry is the gastrointestinal tract

THEN there is suggestive evidence (0.7) that infection is bacterial.

Figure 3: Mycin rule

Another rule-based system (EXPERT) [80], was developed by Weiss. An EXPERT model consisted of hy-potheses (structured as causal and taxonomic networks), findings or observations, and decision rules for logicallyrelating these components. The system was used for rheumatology, ophthalmology [81], and endocrinology. In1983, Peter Politakis [82] developed SEEK system for giving interactive advice about rule refinement duringthe development of an expert system. It was used for dermatology [83].

4

In 80s and 90s, rule-based systems were developed for esophageal cancer [84], psychiatry (Psyxpert: us-ing certainty and importance measures) [85, 86], ascites (using ID3 for rule induction)[87], female urinaryincontinence [88], pregnancy disorders [89], kidney disease [90], metabolic disease [91], nosocomial infections[92], hepato-biliary disorders (HEPAR) [93], neurology [94, 95, 96], appendicitis [97], lung disease [98], thyroiddisorder [99], pathology [100], uterine disease [101] and heart disease [102].

A disadvantage of rule-based applications is that they become increasingly unmanageable when the numberof rules increase [102]. Methods for automated discovery and optimization of the rules have revived interest inthe technique [103, 104]. Rule-based systems are inherently amenable to derivation of decision explanations.

3.4 Heuristic

Heuristic algorithms are those for which there is no formal proof of correctness. These include most of the AIprograms developed in the 1980s and also include rule-based systems which used ad-hoc scoring schemes forinference. CASENET developed by Kulilowski was probably the first heuristic system. It used a descriptivemodel of the disease process for logical interpretation of clinical findings for glaucoma [105]. The causal modelrepresenting pathophysiological mechanisms had the form of a semantic net with weighted links. It was usedto develop consultation systems for neuro-opthalmology, infectious diseases of the eye, rheumatology [106, 107]and pathology, and was extended into the natural language system called CHRONOS [108]. Around the sametime, Graham developed a scoring index for gangrenous and perforating appendicitis [37].

Internist [17] was a wide-domain heuristic system developed for general medicine, at university of Pitts-burgh. It had a knowledge-base of 600 diseases and 4500 findings, developed by a team of physicians andmedical students, who reviewed literature to estimate for each finding, a frequency weight (FW) and evokingstrength (ES) in grade range 1 − 5. FW was the frequency with which the finding was found in a disease andES reflected the belief that a patient had a disease given that the finding was present. In addition, each findinghad an import number (1-5), which reflected the importance of the finding. Internist used a scheme similarto the hypothetico-deductive approach for inference. The medical knowledge-base of INTERNIST was usedto construct CADUCEUS [109] and a commercial system named Quick Medical Reference (QMR) [110] . CA-DUCEUS was designed to infer on cases with simultaneous diseases, and on flawed and scarce data. Inferenceengine of CADUCEUS was similar to MYCIN, but in addition incorporated abductive reasoning.

Other heuristic systems are for diagnosis in geriatric psychiatry (Methuselah) [111], rheumatology [112],cardiology [113] and malformation syndromes [114]. Despite the lack of formal proof of correctness, heuristicmethods have found wide applications. Lack of mathematical rigour lead researchers to explore other modelingtechniques.

3.5 Fuzzy Set theory

Fuzzy sets extend the concept of conventional sets, by allowing its elements to have degrees of membership. Theyare useful for representation of vague concepts expressed in natural language and uncertainties in measurement.The earliest suggestion of the applicability of FST to medicine is ascribed to Zadeh [115]. Early applicationswere largely based on fuzzy relations. However in the last decade, fuzzy rule-based systems have been popularand several other fuzzy set theory based approaches have been developed, as described below.

3.5.1 Fuzzy Relations(FR)

Sanchez modeled medical knowledge as fuzzy relations (figure 4) between symptoms and diseases [116] and usedcompositional rule for inference [117]. Sanchez’s model was extended to develop CADIAG systems. CADIAGknowledge-base was constructed in the form of fuzzy relations acquired from physicians, medical literatureand semi-automatically from patient records. CADIAG captured uncertainty in information about a patient’ssymptoms and medical knowledge, as binary fuzzy relations, and it used a heuristic algorithm for diagnosticinference. These systems were evaluated to have high accuracy for rheumatic [118] and pancreatic [119] diseases.Recently Ciabattoni et al. [120] have defined a formal logical calculus to perform a consistency check of theCADIAG knowledge-base. Applications of fuzzy relations include diagnosis of cardiopathy [121] and arteritis

5

[122]. Recently, Wagholikar et al. evaluated a fuzzy relation (FR) algorithm on a variety of medical datasets[123], and concluded that their algorithm is particularly useful for noisy and incomplete datasets.

‘Probability of finding fever in patient having malaria is 0.8’ translates to

‘Fuzzy occurance relation of fever with malaria is 0.8’

‘Probability of diagnosing malaria in patient who has fever is 0.2’ translates to

‘Fuzzy confirmatory relation of fever with malaria is 0.2’

Figure 4: Fuzzy relation

3.5.2 Fuzzy rules

In these systems, the knowledge-base is generally derived as rule by inducing decision tree from data (figure 5).The crisp rules are then fuzzified by computing fuzzy sets. Such systems have been applied for aphasia [124],sleep apnea [103], coronary disease [125], post-operative infections [126]. hepatic disorder [127], and malaria[128].

If temperature is ‘high’ and lymphocyte count is ‘high’

then confidence for viral infection is 0.95

Figure 5: Fuzzy rule that models linguistic input variables.

3.5.3 Neuro-Fuzzy

Neuro-fuzzy systems combine FST with the theory of ANN. Kuncheva [129] first proposed the model of afuzzy neuron for aviation medicine. Barreto [130] interpreted the connection values and excitation state ofthe neurons. Neuro-fuzzy approach has been applied for diagnosing erythemato-squamous diseases [131], heartdisease [132] and hemorrhage [133].

3.5.4 Other FST based

Artificial Intelligence Research Institute, Spain [134] developed Milord system based on approximate reason-ing for rheumatology (Renoir) and pneumonia (Pneumon-IA). Other FST applications include diagnosis ofpostmenopausal osteoporosis [135], myocardial ischemia [136], hematological disorders [137, 138], sore throat[139], fuzzy-genetic approach [140], fuzzy classification method [141, 142], fuzzy pattern recognition [143], fuzzycognitive maps for pathology [144, 145], fuzzy similarity classifier [146] for erythemato-squamous diseases, in-tutionistic fuzzy sets [147] and fuzzy influence diagrams [148].

FST offers a wide variety of formalisms for modeling medical decision making and is a fertile ground forMDS research [149]. However, the use of fuzzy models generally entails high computational costs.

3.6 Artificial Neural Network

Artificial Neural Network (ANN) is a mathematical model that simulates biological neural networks. It con-sists of an interconnected group of artificial neurons (nerve cells) that process information (figure 6). Thestrength/weight of the neuronal interconnections, topology of the network and choice of excitation function ofthe neurons are important factors governing the behavior and utility of the model. The applications of ANN tomedicine appeared in early 1990s [150] and rapidly gained popularity [151]. ANNs have been applied for cardiol-ogy [152, 153, 154, 155], orthopedics [156], psychiatry [157], low back pain [158], cancer [159, 160, 161, 162], dys-pepsia [163], atrophic gastritis [164], infection [165, 166, 167], Alzheimer’s [168], pulmonary diseases [169, 170],and heat stress [171].

6

However Sargent’s [172] review of literature comparing ANN with standard statistical techniques concludedthat ANN should not replace standard statistical approaches as the method of choice for the classification ofmedical data. This result was supported by studies in urology [173] and oncology [174]. It is suggested thatANN methods should continue to be used and explored with statistical approaches in a complementary manner.A disadvantage of ANN models has been that they do not provide explanations for their decisions. Hence theyare regarded as black-boxes. Recently some progress has been made to address this limitation [175].

Fever

Maliase

Rash

Vomiting

Malaria

Input nodesHidden nodes

Output node

Figure 6: Representation of an artificial neural network (ANN) for diagnosis of malaria with symptoms as inputfeatures.

3.7 Bayesian Network

Bayesian belief networks, also referred to as probabilistic causal networks or Bayesian Networks (BNs), overcomethe assumption of independence between findings which was associated with earlier, NB approach. BNs are amerger of symbolic reasoning and bayesian approaches. The nodes in the network represent the findings andthe links interconnecting the nodes model the inter-dependencies of the findings (figure 7).

The early applications of BNs were published in mid-nineties. BNs have been used for radiology [176],oncology [177, 178, 179, 180], asthma [181], penetrating trauma [182], pneumonia [183], pancreatitis [184],cardiology [185], hematology [186], neuro-muscular disorder [187], and hyper-kinetic disorder [188]. QMR-DT, Pathfinder, Promedas, DIAMED and miniTUBA are some systems using bayesian networks. Shwe et al.[189] reformulated QMR/Internist database into a bayesian network, called QMR-DT. This system showed adiagnostic accuracy comparable to the original Internist I, despite the many approximations that were used inthe conversion process. Promedas is a commercially available system based on knowledge acquired from medicalliterature and medical experts [190]. It includes 2000 diagnoses, 1000 findings and 8600 connections betweendiagnoses and findings, covering a large part of internal medicine. DIAMED has been shown to be useful toreason on medical data [191]. miniTUBA [192] is a web-based modeling system for temporal datasets.

BN models can be complex and practically intractable for some problems. Reducing the complexity of BNand optimization of approximation algorithms for inference, is an active research area [193].

3.8 Support Vector Machines

Support Vector Machine (SVM) methods map data points to a multi-dimensional space and attempt clas-sification by constructing a high dimensional hyperplane to separate the points. One of the earliest re-port using SVM for diagnoses is by Chan [194], for glaucoma detection by identifying patterns obtained byStandard Automated Perimetry (SAP). Utility of SVMs has been reported for diagnosis of cerebral palsy [195],Parkinsonism [196], cancer [197, 198, 199, 200, 201] and diabetic nephropathy [202]. Work on application ofSVM for MDS is at a nascent stage, but with promising results.

7

low serum proteins

ascites

liver disease

Figure 7: Portion of bayesian network for diagnosis of liver disease that models the dependence between ascitesand low serum proteins, both of which can be findings of liver disease.

Heuristic BN

NB

Other FST

Rule-based

SVM

ANN

25%

50%

75%

100%

71-80 81-90 91-00 00-10

Time

Percentage

Figure 8: Distribution of reports covered in this survey over last four decades.

3.9 Other

Apart from the major paradigms of development noted above, several other approaches have been investigatedfor MDS. These include sequential diagnosis [203], cluster analysis [204], discriminant analysis [205, 206, 207],logistic regression [208, 209, 210, 211, 212, 213], partitioning methods [214], Dempster-Schafer theory [215],matrix discrimination [216], modified constellation graph methods [217], degree of similarity [218], nearestneighbor [219], case-based (instance-based or memory-based) learning [220, 221, 222, 223, 224], rough sets[225], decision-tree induction[226, 227, 228], Testor theory [229], information theory [230] and biased min-maxprobability machine [231].

This list of modeling techniques applied for MDS is not exhaustive, and medical diagnosis is a popularproblem for researchers working on classification algorithms. Consequentially, a choice from a wide variety oftechniques is available for MDS application development.

4 Discussion

In this section we have attempted to describe the general trends in research on modeling techniques for MDS,and to identify short-comings as well as provide future directions.

4.1 Lack of decision accuracy

MDS systems for small domains like ECG or EEG classification, have become popular but systems for morecomplex problems have not yet achieved the acceptable degree of accuracy. This continues to be a cardinal

8

obstacle in the utilization of MDS applications in the clinic. There is a need for research to develop accurateclassification algorithms. Apart from the classification/inference algorithm, improved representation of patientinformation may lead to performance improvement. For instance representing temporal and vague concepts[232, 233] and integrating data from diverse sources [234], as elaborated below have been shown to improveaccuracy.

4.2 Semi-automatic to automatic knowledge acquisition

In the 1970s and 80s, many researchers focused on input of physicians to construct the knowledge-base [235,79] for MDS tools. Over the years, the alternative approach of automated analysis of datasets to constructknowledge-base, has become more popular [236, 103]. The latter approach offers several advantages:

• Knowledge-Bases derived from datasets are more precise in comparison with knowledge-bases constructedfrom expert inputs. This is because the inputs provided by human experts are often vague, due to limitedgrades of perception [4, 237].

• When the knowledge-base has a large number of parameters, the process of collating experts’ inputs inthe former approach, becomes cumbersome and practically unfeasible [235, 238]. Automatic approachobviates the logistical difficulties.

• Knowledge-Base constructed using the automated approach captures empirical evidence in the data. Thisapproach aligns with the evidence-based decision making movement which emphasizes on use of empiricalevidence to make clinical decisions [239].

• Medical datasets embed local epidemiological patterns. Hence the derived knowledge-bases can result inmore accurate MDS tools, as disease and symptom patterns vary from one region to another [240, 241,242, 47]. The physician experts on the other hand may not be aware of local trends, especially when theydo not have sufficient experience of clinical practice in a particular locality and when there is a lack ofsystematic studies of disease patterns for the location.

4.3 Choice of modeling technique

The early MDS systems were based on NB model, which made assumptions of conditional independence betweensymptoms and mutual exclusivity of diseases. These systems performed as-well-as, and at times better thanexperts. However, despite their performance NB did not receive wider applications, for the following reasons[243]: 1) the assumptions of NB, were considered as an over-simplification and 2) the quantitative approach ofBayes rule was perceived as inappropriate representation of the qualitative way in which humans reason.

AI techniques which were emerging at that time, appeared to provide alternative solution to bypass theseproblems. The Confidence Factor (CF) [244] and QMR inference model, used heuristic methods for inferenceand instead focused on representing large amount of expert knowledge. However, the heuristic inference mech-anisms did not avoid the assumptions of conditional independence in NB; they merely obscured them [245].Theoretical studies have shown that these heuristic schemes and the Dempster Schafer [246] model were relatedto odds likelihood updating scheme [247, 248]. An additional disadvantage of the AI systems was that the pa-rameters for these systems were more likely to be erroneous as compared to the NB model. This is because theassessment of symptom-disease relationships for these systems required the physicians to provide judgements interms of likelihood of a disease given a finding in a direction that is considered cognitively challenging [249, 243].In contrast, the use of Bayes rule required the estimation of probabilities in terms of likelihood of a symptomgiven a disease, which is natural to physicians. In an empirical study, Heckerman [250] found that NB hadbetter diagnostic accuracy than CF.

These developments led researchers in 1990s to revert to the principled probability based approach forinference, but now coupling it with expressive and richer knowledge representation schemes, which led to thedevelopment of BNs. Other methods like SVM, ANN and Information theory have also been explored. However,despite the development of new complex models [174], no significant improvement in accuracy has been achievedand many comparative studies have evaluated NB as near optimal [251, 39, 40]. NB remains the most widely

9

researched approach and many researchers regard NB as the reference standard and use it as a benchmark forperformance of any new algorithm. The soft computing methods including FST, Evolutionary Algorithm (EA),ANN, and BN are computationally intensive. These methods have benefitted from the availability of cheapcomputational power that has rendered the feasibility of some application, earlier regarded as intractable.

Overall a wide choice of modeling techniques is available for developers of MDS tools. However, there isa dearth of studies which compare the performance of different techniques [252]. Hence, we suggest that adiversity of techniques are evaluated for the problem under consideration, in order to decide on the optimaltechnique.

4.4 Integrating clinical and genomic patient information

One of the reasons, why MDS tools may be found to have sub-optimal accuracy is that the training data mayitself lack all the attributes that are required for decision making. Combining decision support methodologiesthat process information stored in different data formats has been shown to improve MDS performance [253].Apart from laboratory information, attributes extracted from free text information, and even gene profiling datacould be combined in the Electronic Medical Record (EMR). Decision models built from such an integratedEMR may possibly lead to clinically applicable systems. However gene profiles are complex and analysis of suchdata is currently limited to bioinformatics applications [254].

4.5 Modeling temporal information

Knowledge about the order of occurrence of events, has a significant bearing on clinical decision. For example,head injury may occur after fainting or fainting may occur after head injury – these situations differ in thecausative factor and consequently the diagnosis. Currently, most MDS systems are not designed with a formalrepresentation of time, and they are unable to reason about temporal relations of events. In the last decade, sometemporal models [255, 256] have been developed and there exists empirical evidence that temporal modelingcan improve MDS accuracy [257, 145]. Temporal modeling of medical information is challenging and may holdthe key for improving performance of MDS systems. The reader is referred to [258, 259, 260] for further reading.

4.6 Modeling fuzziness

Probability based approaches have dominated MDS research. However a problem with probabilistic approachesis that they do not differentiate between the types of uncertainties, and model all uncertainty using probabilitytheory [15]. Uncertainty in the entities involved in medical decision making can be differentiated into ignorance,non-specificity, vagueness and strife [261]. FST provides a promising alternative, as it offers both– a mechanismto model the different types of uncertainties as well as a principled approach for inference using fuzzy entities.Several studies have demonstrated the utility of fuzzy logic concepts in medical domain (section 3.5). FST hasbeen recently of active interest for MDS research.

FST can effectively model medical reasoning because, symptom and disease entities in medicine are definedusing linguistic descriptions which are inherently vague/fuzzy. The imprecise definitions along with the humancognitive limitation render the physicians reasoning process vague and subjective. Although models based onFST generally provide a less precise representation of a problem, as compared to models based on Aristotelianlogic (including probability based models), FST based models have been known to closely approximate humanintelligence [261]. Hence, the FST models could provide a better approximation of physician’s diagnosticprocess, which is characterized by the lack of precision in knowledge and an inference mechanism, which haseluded attempts of probabilistic modelers.

4.7 Evaluation measures

MDS systems are often designed to output differential diagnosis. This is essentially a list of decision classesin the decreasing order of their confidence scores. Medical diagnostic problems often involve more than twodecision classes (referred to as multi-class problem), which complicates the computation of decision accuracy ofthe differential diagnosis. This is because much of the work on evaluation measures for classification algorithms,

10

has been carried out for problems in which the decision attribute is binary– there are only two decision classes.Most of the evaluation studies carried out before the last two decades, measured performance by identifyingthe rank of the correct diagnosis in the list of differential diagnosis– lower ranks meant better accuracy. Acut-off rank was used, and the percentage of times the correct diagnosis was present below the cut-off rank, wasreported as the accuracy. However this measure is subject to the number of diagnoses considered in the systemand also does reflect the sensitivity and specificity of the system. To address this problem, some researchersstarted reporting area under Receiver Operating Characteristic (ROC) [262], for binary classification problems.Recently measures generalizing the computation of ROC for multi-class problems have also been developed[263]. However, researchers continue to report evaluation results using different measures, which impedes thetask of comparing the algorithms. Further research is needed to establish standard evaluation measures forreporting performance of Medical Decision Support System (MDS) applications.

4.8 Evaluation in the clinical setting

MDS models are usually evaluated on small or simulated datasets for preliminary studies, followed by compre-hensive evaluations on medium to large (> 200)[172] real datasets (see figure 9). Few evaluation studies areperformed on large datasets [236, 264]. Reliability of the results generally increases with size of the datasetand number of evaluative studies. Evaluations are usually retrospective, in which the case records of previouslylabelled patients are used as gold standard. Rarely, MDS models are evaluated prospectively in the clinicalsetting, where the model’s decision are compared with those of the physician in real-time. The latter evalua-tions are most useful to build confidence about computer based decision aids in the medical community. MDSresearchers should perform retrospective evaluations first, and when their algorithms are found to be sufficientlyaccurate, should aim for prospective studies [150, 265]. De Dombal pioneered clinical evaluations of MDS sys-tems. His Leeds system was evaluated first retrospectively on past patient records, and then with real timeprospective studies across many locations involving practicing surgeons [38]. This research played a significantrole in impressing the practical utility of MDS systems.

Dataset size

Fre

quen

cy

0 200 400 600 800 1000 1200

05

1525

Figure 9: Histogram of the number of data instances in small and medium-large datasets investigated in thesurveyed literature

4.9 Complexity of decision problems

The complexity of a decision problem depends on i) domain, ii) number of instances in training data, iii) numberand type of attributes, and iv) number of diagnostic labels considered. Medical domains vary in the degree ofcomplexity of the classification problem they deal with. For instance, cardiovascular diseases which occur withnumerous symptoms, present a low complexity problem as compared to cancers which are insidious with few

11

symptoms. In addition, a decision problem may be complex due to dearth of training instances and informationattributes. Binary diagnostic problems in which the presence or absence of a disease is to be computed, are lesscomplex than problems in which patients can have one or more diagnosis from a diagnosis set. The problemdomains addressed in MDS research are diverse (see table 1 and figures 10 and 9). Researchers should evaluatesimpler problems for preliminary studies and should progress towards increasingly complex problems to explorethe utility of their algorithms.

Cardiovascular

Endocrinology

Gastroenterology

Oncology

Hematology

Immunology

Neurology

Other

Psychiatry

Respiratory

Figure 10: Distribution of studies cited in this paper, across medical specialities.

4.10 Decision Utility

After arriving at a provisional diagnosis based on patient interview, physicians consider severity of the dif-ferential diagnoses and individual patient preferences, to calculate the utility of diagnoses or expected utilityof prescribing additional tests [239, 78]. Utility considerations involve the risk to the patient and cost of thetests (see section 3.2). Physicians have a tendency to give a guarded opinion [280], wherein they do not easilydiscard diagnosis, which poses potential risks to the patient. The lack of such utility considerations, could bea reason why evaluations comparing the decisions of MDS tools with physician’s differential diagnosis, reportlow accuracy. This factor is expected to play a lesser role, when the final diagnosis arrived at discharge of thepatient from the health care facility, is considered as the gold standard for measuring performance accuracy.Hence, research is needed to investigate the significance of decision utility for MDS evaluation experiments.

4.11 Utilization of High Performance Computing

Technological advances in the field of high performance [281] computing have made cheap computational poweravailable to individuals. Multi-core processors and computers in parallel configurations have been constructedfor high end computational applications. This provides an opportunity for exploring MDS modeling techniquesthat were regarded practically intractable, a few decades ago. MDS researchers should harness these resources,so developed MDS tools for increasingly complex medical decision problems.

4.12 Opensource

Many MDS systems are developed for commercial applications. Their decision algorithms or knowledge-basesare rarely released into public domain. The medical datasets used for evaluations are usually withheld. Non-availability of the tools, knowledge-bases and data impedes replication of the evaluation experiments as well ascomparative studies. If researchers were to release these materials in opensource, it may be possible to reusethem in evaluation experiments of other algorithms, and useful benchmarks can be obtained. The importanceof the need to release the tools and datasets as supplementary material with the reports, was realized early onby researchers in bioinformatics and genomic domain. Consolidation of their efforts has fostered advances inthose fields. A similar practice needs to be encouraged in MDS research. Another reason why MDS tools needto be opensource is their potential application in the clinical setting, where error-free operation is called for[282].

12

Table 1: Diagnostic problems addressed in surveyed literature

Cardiovascular myocardial-infarction [150, 69, 208, 152, 210], coronary disease [215, 125], arrhythmia [104], car-diac disease [113, 121, 185, 231, 102, 132, 154], giant cell arterits [153], hypertension [68], pulmonaryembolism [266], valvular disease [155]

ENT otoneurological disease [227], apnea [103], sore throat [139]

Gastroenterology hepatobiliary [127, 216, 93], appendicitis [166, 37, 41, 37], bowel disease [71], obstruction [72],cirrhosis [267, 49], dyspepsia [163, 226], hepatitis [146, 123, 47, 51, 48, 50], bleeding [73, 133], painin abdomen [123, 38, 27, 251, 40, 42, 55], pancreatitis [184, 119, 53], ascites [87],

Neurology Alzheimer’s [168], Parkinsonism [196], acute event [94], aphasia [124], cerebral palsy [195],epilepsy [60], headache/facial pain [96], general [268], neuromuscular dissorders [187, 269]

Oncology cancer of breast [161, 161, 179, 142, 146, 123, 231, 222, 198, 201], cerebrum [46], cervix [270, 271, 178],esophagus [84], stomach [164], ovary [197], prostate [173], skin [162, 83], lung [45, 54]; leukemia[159, 160, 141], melanoma [199], neuroma [44], general [43, 200]

Psychiatry ADHD [188], geriatric [111], general [157, 206, 85, 86]

Reproductive contraception [123], hysterectomy [101], gynecological pain in abdomen [39], pregnancy dis-orders [89], vaginal discharge [59]

Respiratory asthma [181, 212], interstitial disease [32, 209], lung disease [98, 169], pneumonia [169, 170, 183, 134],pleural effusion [75],

Endocrinology diabetes [146, 224], thyroid disorder [272, 146, 99], general [268]

Hematology anaemia [138, 67], blood donation [123], general [186, 36, 66, 273]

Immunology rheumatic disease [134, 118, 107, 112, 61, 62, 221, 82]

Infectious Lyme’s [167], MSRA [165], malaria [128], post-operative infection [126], nosocomial infection[92]

Musculoskeletal low-back pain [158], arthritis [122], intrabony lesion [57], osetoporosis [135]

Opthalmology glaucoma [81, 194], leukocoria [65], general [268]

Other penetrating trauma [182], pediatric [230], critical care [274] disease outbreak [31], triage [275],pulpal pain [56], renal [90, 276], urinary incontinence [88], pathology [277, 144], general [190,

123, 110, 17, 28, 80, 191, 25], radiology [176, 177, 278, 279], heat-stress [171], metabolic disorder [91],congenital anomaly [33, 34, 114], adverse drug reaction [70], lypmh-node [78], rare syndromes[220], skin disease [64, 63, 146, 131]

13

4.13 Using free-text data

A large proportion of patient information is in the free text form. This is mainly comprised of a verbatim reportof patient chief-complaints [283] and physicians’ notes. Such data lacks structure and is not amenable to analysis.Hence, it is largely unused by decision support tools. For extracting information attributes, free-text is parsed byNLP algorithms to identify the parts of speech and these are mapped to a standard vocabulary (ontology). Thisarea of research is rapidly progressing [284] and applications [285, 21, 286, 287] based on free-text parsing havebeen developed. With advances in this area, free-text can provide useful information attributes, and integrationof these attributes with laboratory-test based ones will enhance the decision accuracy of MDS applications.

Unified Medical Language System (UMLS) [288] reflects the progress in representing complex medical con-cepts using ontologies. Advances in ontology and parsing would hugely facilitate the extraction of features fromfree text data, that was not earlier amenable to analysis. All these developments, along with the increased useof Electronic Medical Record (EMR) systems in hospitals, can be expected to lead to the availability of hugeamounts of data features. MDS applications would be needed to harness the intelligence derieved from data inthe EMR systems.

5 Concluding Remarks

A primary goal of biomedical informatics is to provide computer based decision aids to physicians. Research fordeveloping increasingly accurate MDS algorithms is necessary to achieve this goal. The need for MDS can beexpected to become increasingly acute, as information explosion in medical science continues. Implementationof EMR systems at health-care centers can facilitate adoption of MDS applications.

The authors suggest that – i) Improvement in the accuracy of MDS application may be possible by mod-eling of vague and temporal data, research on inference algorithms, integration of patient information fromdiverse sources and improvement in gene profiling algorithms; ii) MDS research would be facilitated by publicrelease of de-identified medical datasets, and development of opensource data-mining tool kits; iii) Comparativeevaluations of different modeling techniques are required to understand characteristics of the techniques and toguide developers in the choice of technique for a particular medical decision problem; iv) Evaluations of MDSapplications in clinical setting are necessary to foster physicians’ utilization of these decision aids.

Acronyms

AI Artificial Intelligence.

ANN Artificial Neural Network.

BN Bayesian Network.

CF Confidence Factor.

EA Evolutionary Algorithm.

ECG Electrocardiogram.

EMR Electronic Medical Record.

FR fuzzy relation.

FST fuzzy set theory.

MDS Medical Decision Support System.

14

MDS Medical Decision Support.

NB naive bayes.

NLP Natural Language Processing.

QMR Quick Medical Reference.

ROC Receiver Operating Characteristic.

SAP Standard Automated Perimetry.

SVM Support Vector Machine.

UMLS Unified Medical Language System.

6 Acknowledgments

The authors are grateful for the excellent suggestions and comments of the reviewers that helped improve themanuscript. The first author was supported by a fellowship from Indian Council of Medical Research (ICMR),when part of this research was carried out.

References

[1] E. S. Berner, R. S. Maisiak, C. G. Cobbs, and O. D. Taunton. Effects of a decision support system on physicians’diagnostic performance. Journal of the American Medical Informatics Association : JAMIA, 6(5):420–427, 1999.

[2] A. Hall and G. Walton. Information overload within the health care system: a literature review. Health Info LibrJ, 21(2):102–108, June 2004.

[3] J. Wyatt. Use and sources of medical knowledge. Lancet, 338(8779):1368–1373, November 1991.

[4] G. A. Miller. The magical number seven plus or minus two: some limits on our capacity for processing information.Psychol Rev, 63(2):81–97, 1956.

[5] R. Glassman, K. Leniek, and T. Haegerich. Human working memory capacity is 7 2 in a radial maze withdistracting interruption: possible implication for neural mechanisms of declarative and implicit long-term memory.Brain Research Bulletin, 47(3):249–256, October 1998.

[6] P. D. Clayton, R. S. Evans, T. Pryor, R. M. Gardner, P. J. Haug, O. B. Wigertz, and H. R. Warner. Bringing helpto the clinical laboratory–use of an expert system to provide automatic interpretation of laboratory data. Annalsof clinical biochemistry, 24 Suppl 1:5–11, 1987.

[7] Richard Smith. What clinical information do doctors need? BMJ, 313(7064):1062–1068, October 1996.

[8] Jon Brassey, Glyn Elwyn, Chris Price, and Paul Kinnersley. Just in time information for clinicians: a questionnaireevaluation of the attract project. BMJ, 322(7285):529–530, March 2001.

[9] Gordon D. Schiff, Seijeoung Kim, Richard Abrams, Karen Cosby, Bruce Lambert, Arthur S. Elstein, Scott Hasler,Nela Krosnjar, Richard Odwazny, Mary F. Wisniewski, and Robert A. McNutt. Diagnosing Diagnosis Errors:Lessons from a Multi-institutional Collaborative Project. Advances in Patient Safety 2005;2:255-278., volume 2,pages 255–278. 2005.

[10] Mark L. Graber, Nancy Franklin, and Ruthanna Gordon. Diagnostic error in internal medicine. Arch Intern Med,165(13):1493–1499, July 2005.

[11] David B. Aronow, Thomas H. Payne, and S. Pierre Pincetl. Postdoctoral training in medical informatics: A surveyof national library of medicine-supported fellows. Med Decis Making, 11(1):29–32, February 1991.

[12] Robert S. Ledley and Lee B. Lusted. Reasoning foundations of medical diagnosis; symbolic logic, probability, andvalue theory aid our understanding of how physicians reason. Science, 130(3366):9–21, 1959.

15

[13] Randolph A. Miller. Medical diagnostic decision support systemspast, present, and future. Journal of the AmericanMedical Informatics Association, 1(1):8–27, January 1994.

[14] C. A. Kulikowski. Artificial intelligence in medicine: a personal retrospective on its emergence and early function.In Proceedings of ACM conference on History of medical informatics, New York, NY, USA, 1987. ACM Press.

[15] F. T. De Dombal. Assigning value to clinical information–a major limiting factor in the implementation of decision-support systems. Methods of information in medicine, 35(1):1–4, March 1996.

[16] B. Puppe, C. Ohmann, K. Goos, F. Puppe, and O. Mootz. Evaluating four diagnostic methods with acute abdominalpain cases. Methods Inf Med, 34(4):361–368, September 1995.

[17] J. D. Myers. The computer as a diagnostic consultant, with emphasis on use of laboratory data. Clin Chem,32(9):1714–1718, September 1986.

[18] E. H. Shortliffe. Medical thinking: What should we do?, June 2006.

[19] Elaine F. Chouinard, Bernard L. Ryack, and Douglas M. Stetson. A comparison of the usability of three versionsof a computerized medical diagnostic assistance program for abdominal pain. Technical report, Naval SubmarineMedical Research Lab., Groton ,CT, July 1991.

[20] D. W. Bates. Ten commandments for effective clinical decision support: Making the practice of evidence-basedmedicine a reality. Journal of the American Medical Informatics Association, 10(6):523–530, August 2003.

[21] Carol Friedman, Lyudmila Shagina, Yves Lussier, and George Hripcsak. Automated encoding of clinical documentsbased on natural language processing. Journal of the American Medical Informatics Association : JAMIA, 11(5):392–402, 2004.

[22] Jonathan Clive, Lee B. Lusted, Casimir Kulikowski, T. Allan Pryor, and Bruce McCormick. Computer algorithmsfor analyzing patient data. In AFIPS ’74: Proceedings of the May 6-10, 1974, national computer conference andexposition, page 1025, New York, NY, USA, 1974. ACM.

[23] J. Ridderikhoff and B. van Herk. Who is afraid of the system? doctors’ attitude towards diagnostic systems.International journal of medical informatics, 53(1):91–100, January 1999.

[24] H. R. Warner, A. F. Toronto, L. G. Veasey, and R. Stephenson. A mathematical approach to medical diagnosis.application to congenital heart disease. JAMA : the journal of the American Medical Association, 177:177–183, July1961.

[25] R. S. Evans. The help system: a review of clinical applications in infectious diseases and antibiotic use. M.D.computing : computers in medical practice, 8(5), 1991.

[26] Robert M. Kolodner and J. V. Douglas, editors. Computerizing Large Integrated Health Networks: the Va Success.Springer, 1 edition, February 1997.

[27] F. T. de Dombal. Computers, diagnoses and patients with acute abdominal pain. Archives of emergency medicine,9(3):267–270, September 1992.

[28] G. O. Barnett, J. J. Cimino, J. A. Hupp, and E. P. Hoffer. Dxplain. an evolving diagnostic decision-support system.JAMA : the journal of the American Medical Association, 258(1):67–74, July 1987.

[29] A. Gammerman and A. R. Thatcher. Bayesian diagnostic probabilities without assuming independence of symptoms.Methods of information in medicine, 30(1):15–22, 1991.

[30] I. Zelic, I. Kononenko, N. Lavrac, and V. Vuga. Induction of decision trees and bayesian classification applied todiagnosis of sport injuries. Journal of medical systems, 21(6):429–444, December 1997.

[31] Tom Burr, Frederick Koster, Rick Picard, Dave Forslund, Doug Wokoun, Ed Joyce, Judith Brillman, Phil Froman,and Jack Lee. Computer-aided diagnosis with potential application to rapid detection of disease outbreaks. Statisticsin medicine, 26(8):1857–1874, April 2007.

[32] I. J. Check, G. T. Gowitt, and G. W. Staton. Bronchoalveolar lavage cell differential in the diagnosis of sarcoidinterstitial lung disease. likelihood ratios based on computerized data base. American journal of clinical pathology,84(6):744–747, December 1985.

[33] D. F. Schorderet. Diagnosing human malformation patterns with a microcomputer: evaluation of two differentalgorithms. American journal of medical genetics, 28(2):337–344, October 1987.

[34] F. Wiener, M. Gabbai, and M. Jaffe. Computerized classification of congenital malformations using a modifiedbayesian approach. Computers in biology and medicine, 17(4):259–267, 1987.

16

[35] C. Y. Lee, L. Carmony, M. Evens, F. Naeymi-Rad, and D. Trace. A test selection module for medas. In Symposiumon Computer Applications in Medical Care, pages 706–710, 1991.

[36] R. L. Engle, B. J. Flehinger, S. Allen, R. Friedman, M. Lipkin, B. J. Davis, and L. L. Leveridge. Heme: a computeraid to diagnosis of hematologic disease. Bulletin of the New York Academy of Medicine, 52(5):584–600, June 1976.

[37] D. F. Graham. Computer-aided prediction of gangrenous and perforating appendicitis. British medical journal,2(6099):1375–1377, November 1977.

[38] I. D. Adams, M. Chan, P. C. Clifford, W. M. Cooke, V. Dallos, F. T. de Dombal, M. H. Edwards, D. M. Hancock,D. J. Hewett, and N. McIntyre. Computer aided diagnosis of acute abdominal pain: a multicentre study. Br Med J(Clin Res Ed), 293(6550):800–804, September 1986.

[39] B. S. Todd and R. Stamper. The relative accuracy of a variety of medical diagnostic programs. Methods Inf Med,33(4):402–416, October 1994.

[40] C. Ohmann, V. Moustakis, Q. Yang, and K. Lang. Evaluation of automatic knowledge acquisition techniques in thediagnosis of acute abdominal pain. acute abdominal pain study group. Artif Intell Med, 8(1):23–36, February 1996.

[41] F. H. Edwards and R. S. Davies. Use of a bayesian algorithm in the computer-assisted diagnosis of appendicitis.Surgery, gynecology & obstetrics, 158(3):219–222, March 1984.

[42] D. H. Wilson, P. D. Wilson, R. G. Walmsley, J. C. Horrocks, and F. T. De Dombal. Diagnosis of acute abdominalpain in the accident and emergency department. The British journal of surgery, 64(4):250–254, April 1977.

[43] F. Wiener, D. Laufer, and A. Ribak. Computer-aided diagnosis of odontogenic lesions. International journal of oraland maxillofacial surgery, 15(5):592–596, October 1986.

[44] J. R. Iglesias, J. Esparza, C. Aruffo, and K. Maier-Hauff. Differential diagnosis of intraspinal neurinomas andmeningiomas by means of a bayesian system. Archivos de neurobiologia, 51(6):333–341, 1988.

[45] F. H. Edwards, P. S. Schaefer, S. Callahan, G. M. Graeber, and R. A. Albus. Bayesian statistical theory in thepreoperative diagnosis of pulmonary lesions. Chest, 92(5):888–891, November 1987.

[46] G. H. Du Boulay, D. Teather, D. Harling, and G. Clarke. Improvement in the computer-assisted diagnosis of cerebraltumours. The British journal of radiology, 50(600):849–854, December 1977.

[47] G. Lindberg, A. Bjorkman, and R. P. Knill-Jones. Computer aided diagnosis of jaundice. a comparison of two databases. Scandinavian journal of gastroenterology. Supplement, 128:180–189, 1987.

[48] Y. Reisman, G. M. van Dam, C. H. Gips, S. M. Lavelle, B. Kanagaratnam, P. Niermeijer, P. Spoelstra, andO. de Vries. Physician’s working diagnosis compared to the euricterus real life data diagnostic tool trial in three jaun-dice databases: Euricterus dutch, independent prospective and independent retrospective. Hepato-gastroenterology,44(17):1367–1375, 1997.

[49] Y. Reisman, C. H. Gips, and S. M. Lavelle. Primary biliary cirrhosis: an electronic diagnostic tool trial basedon symptoms, (past) history and signs only, using the european database euricterus. the euricterus pmg. Hepato-gastroenterology, 44(16):1104–1109, 1997.

[50] S. M. Lavelle and B. Kanagaratnam. The information value of clinical data. International journal of bio-medicalcomputing, 26(3):203–209, September 1990.

[51] A. Malchow-Møller, C. Thomsen, P. Matzen, L. Mindeholm, B. Bjerregaard, S. Bryant, J. Hilden, J. Holst-Christensen, T. S. Johansen, and E. Juhl. Computer diagnosis in jaundice. bayes’ rule founded on 1002 consecutivecases. Journal of hepatology, 3(2):154–163, 1986.

[52] F. Begon, A. M. Lockhart, J. M. Metreau, and D. Dhumeaux. A computer-aided system for the diagnosis of hepato-biliary diseases. a comparison with the performance of physicians. Medical informatics = Medecine et informatique,4(1):35–42, 1979.

[53] M. de Bernardinis, V. Violi, L. Roncoroni, M. Montanari, and A. Peracchia. Automated selection of high-riskpatients with acute pancreatitis. Critical care medicine, 17(4):318–322, April 1989.

[54] J. Polak, J. Polak, and A. Kubık. [the bayesian statistical theory in the diagnosis of malignant and non-malignantdiseases of the lung, pleura and mediastinum]. Casopıs lekaru ceskych, 132(20):609–615, October 1993.

[55] J. Wojtowicz and W. Adamczyk. A failure to improve radiologists’ performances in diagnosing pulmonary lesionsby a computer-aided approach. RoFo : Fortschritte auf dem Gebiete der Rontgenstrahlen und der Nuklearmedizin,123(1):10–12, July 1975.

17

[56] B. D. Monteith. Computerized expert system for the diagnosis of pulp-related pain. The International journal ofprosthodontics, 4(1):30–36, 1991.

[57] S. C. White. Computer-aided differential diagnosis of oral radiographic lesions. Dento maxillo facial radiology,18(2):53–59, May 1989.

[58] T. Chard. Qualitative probability versus quantitative probability in clinical diagnosis: a study using a computersimulation. Medical decision making : an international journal of the Society for Medical Decision Making, 11(1):38–41, 1991.

[59] T. Chard. Human versus machine: a comparison of a computer ’expert system’ with human experts in the diagnosisof vaginal discharge. International journal of bio-medical computing, 20(1-2):71–78, January 1987.

[60] N. I. Kaliadin, N. P. Kulikova, V. T. Lekomtseva, A. A. Suntsov, and A. N. Sheinin. [the diagnosis of epilepsy inthe interactive mode using computer technology]. Meditsinskaia tekhnika, (3):40–42, 1996.

[61] H. J. Bernelot Moens and J. K. van der Korst. Comparison of rheumatological diagnoses by a bayesian programand by physicians. Methods of information in medicine, 30(3):187–193, August 1991.

[62] H. J. Moens and J. K. van der Korst. Development and validation of a computer program using bayes’s theorem tosupport diagnosis of rheumatic disorders. Annals of the rheumatic diseases, 51(2):266–271, February 1992.

[63] H. Kolles and K. Remberger. How to build a computer-assisted, diagnosis-finding system. an example in der-matopathology. Archives of pathology & laboratory medicine, 115(10):1011–1015, October 1991.

[64] G. J. Brooks, R. E. Ashton, and R. J. Pethybridge. Dermis: a computer system for assisting primary-care physicianswith dermatological diagnosis. The British journal of dermatology, 127(6):614–619, December 1992.

[65] A. S. Leveille, K. J. Fritz, W. M. Jay, and S. J. Silverman. Bayes’ theorem in ophthalmologic computer diagnosis.Journal of pediatric ophthalmology and strabismus, 19(2):94–96, 1982.

[66] D. T. Nguyen, L. W. Diamond, G. Priolet, and C. Sultan. Expert system design in hematology diagnosis. Methodsof information in medicine, 31(2):82–89, June 1992.

[67] F. Sigaux, M. Imbert, G. Priolet, J. J. Bucquen, C. Levy, and C. Sultan. [an aid in decision-making in hematology:characteristics and performances of the program. 200 cases of anemia]. Presse medicale (Paris, France : 1983),16(3):111–114, January 1987.

[68] A. Blinowska, G. Chatellier, A. Wojtasik, and J. Bernier. Diagnostica–a bayesian decision-aid system–applied tohypertension diagnosis. IEEE transactions on bio-medical engineering, 40(3):230–236, March 1993.

[69] K. S. Bay, S. J. Lee, D. P. Flathman, J. W. Roll, and W. Piercy. Application of step-wise discriminant analysisand bayesian classification procedure in determining prognosis of acute myocardial infarction. Canadian MedicalAssociation journal, 115(9):887–892, November 1976.

[70] K. L. Lanctot and C. A. Naranjo. Computer-assisted evaluation of adverse events using a bayesian approach. Journalof clinical pharmacology, 34(2):142–147, February 1994.

[71] W. D. Jeans and A. F. Morris. The accuracy of radiological and computer dianoses in small bowel examinations inchildren. The British journal of radiology, 49(584):665–669, August 1976.

[72] A. Bogusevicius, J. Pundzius, A. Maleckas, and L. Vilkauskas. Computer-aided diagnosis of the character of bowelobstruction. International surgery, 84(3):225–228, 1999.

[73] C. Ohmann, M. Kunneke, R. Zaczyk, K. Thon, and W. Lorenz. Selection of variables using ’independence bayes’ incomputer-aided diagnosis of upper gastrointestinal bleeding. Statistics in medicine, 5(5):503–515, 1986.

[74] W. B. Schwartz, G. A. Gorry, J. P. Kassirer, and A. Essig. Decision analysis and clinical judgment. The Americanjournal of medicine, 55(3):459–472, October 1973.

[75] A. S. Ginsberg. Decision Analysis in Clinical Patient Management with an Application to the Pleural EffusionProblem. PhD thesis, Dept. of Engineering -Economic systems, Stanford University, January 1970.

[76] G. A. Gorry, J. P. Kassirer, A. Essig, and W. B. Schwartz. Decision analysis as the basis for computer-aidedmanagement of acute renal failure. The American journal of medicine, 55(3):473–484, October 1973.

[77] P. Szolovits, R. S. Patil, and W. B. Schwartz. Artificial intelligence in medical diagnosis. Annals of internal medicine,108(1):80–87, January 1988.

[78] D. E. Heckerman, E. J. Horvitz, and B. N. Nathwani. Toward normative expert systems: Part i. the pathfinderproject. Methods of information in medicine, 31(2):90–105, June 1992.

18

[79] Edward H. Shortliffe. Mycin: A knowledge-based computer program applied to infectious diseases. In Annu SympComput Appl Med Care, pages 66–69, October 1977.

[80] R. N. Goldberg and S. M. Weiss. An experimental transformation of a large expert knowledge base. Journal ofmedical systems, 6(1):41–52, February 1982.

[81] G. A. Drastal and C. A. Kulikowski. Knowledge-based acquisition of rules for medical diagnosis. Journal of medicalsystems, 6(5):433–445, October 1982.

[82] Peter G. Politakis. Using empirical analysis to refine expert system knowledge bases (seek). PhD thesis, NewBrunswick, NJ, USA, 1983.

[83] A. D. Vanker and W. Van Stoecker. An expert diagnostic program for dermatology. Computers and biomedicalresearch, an international journal, 17(3):241–247, June 1984.

[84] S. I. Danilenko, A. F. Chernousov, A. P. Pozdniakov, V. I. Fokin, and E. F. Stranadko. [differential diagnosis ofesophageal cancer by using mathematical decision rules]. Voprosy onkologii, 25(7):22–25, 1979.

[85] M. A. Overby. Psyxpert: an expert system prototype for aiding psychiatrists in the diagnosis of psychotic disorders.Computers in biology and medicine, 17(6):383–393, 1987.

[86] R. T. Plant, S. Murrell, and H. R. Moreno. Prototype decision support system for a differential diagnosis of psychotic,mood, and organic mental disorders: Part ii. Medical decision making : an international journal of the Society forMedical Decision Making, 14(3):273–288, 1994.

[87] E. L. Kinney, D. Brafman, and R. J. Wright. An expert system on the diagnosis of ascites. Computers and biomedicalresearch, an international journal, 21(2):169–173, April 1988.

[88] P. A. Riss and H. Koelbl. Development of an expert system for preoperative assessment of female urinary inconti-nence. International journal of bio-medical computing, 22(3-4):217–223, 1988.

[89] J. Ruszkowski. Early pregnancy disorders: expert knowledge based consultation. Journal of perinatal medicine,16(4):289–297, 1988.

[90] M. Ivandic, Y. Ogurol, W. Hofmann, and W. G. Guder. From a urinalysis strategy to an evaluated urine proteinexpert system. Methods of information in medicine, 39(1):93–98, March 2000.

[91] R. Hofestadt. A rule based system for the detection of metabolic diseases. Medinfo. MEDINFO, 8 Pt 2:964–968,1995.

[92] J. Joch, T. Burkle, and J. Dudeck. Decision support for infectious diseases–a working prototype. Studies in healthtechnology and informatics, 77:812–816, 2000.

[93] P. J. Lucas, R. W. Segaar, and A. R. Janssens. Hepar: an expert system for the diagnosis of disorders of the liverand biliary tract. Liver, 9(5):266–275, October 1989.

[94] R. Cavestri, L. Radice, V. D’Angelo, and E. Longhini. Focus. an expert system for the clinical diagnosis of thelocation of acute neurologic events. Minerva medica, 82(12):815–820, December 1991.

[95] G. Rom, G. Schwarz, R. Grims, E. Rumpl, G. Pfurtscheller, and V. Haase. Braindex: an interactive, knowledge-based system supporting brain death diagnosis. Methods of information in medicine, 29(3):193–199, July 1990.

[96] Y. Matsumura. Rhinos: a consultation system for diagnosis of headache and facial pain. Computer methods andprograms in biomedicine, 23(1):65–71, August 1986.

[97] R. N. Shiffman and R. A. Greenes. Use of augmented decision tables to convert probabilistic data into clinicalalgorithms for the diagnosis of appendicitis. In Annual Symposium on Computer Application in Medical Care, pages686–690, 1991.

[98] A. Kar, G. E. Miller, and S. V. Sheppard. Pulmonologist: a computer-based diagnosis system for pulmonary diseases.International journal of bio-medical computing, 21(3-4):223–235, November 1987.

[99] P. E. File, P. I. Dugard, and A. S. Houston. Evaluation of the use of induction in the development of a medicalexpert system. Computers and biomedical research, an international journal, 27(5):383–395, October 1994.

[100] J. P. Baak and P. H. Kurver. Development and use of a rule-based pathology expert consultation system. Analyticaland quantitative cytology and histology / the International Academy of Cytology [and] American Society of Cytology,10(3):214–218, June 1988.

[101] R. Varma and V. Chankong. Computer-aided decisions for prospective hysterectomy screening. Health matrix,7(3):30–32, 1989.

19

[102] Carlos Ordonez. Association rule discovery with the train and test approach for heart disease prediction. IEEEtransactions on information technology in biomedicine : a publication of the IEEE Engineering in Medicine andBiology Society, 10(2):334–343, April 2006.

[103] Mila Kwiatkowska, Stella S. Atkins, Najib T. Ayas, and Frank F. Ryan. Knowledge-based data analysis: first steptoward the creation of clinical prediction rules using a new typicality measure. IEEE transactions on informationtechnology in biomedicine, 11(6):651–660, November 2007.

[104] Markos G. Tsipouras, Costas Voglis, and Dimitrios I. Fotiadis. A framework for fuzzy expert system creation–application to cardiovascular diseases. IEEE transactions on bio-medical engineering, 54(11):2089–2105, November2007.

[105] Saul Amarel. Introduction to the comtex microfiche edition of the rutgers university artificial intelligence researchreports: The history of artificial intelligence at rutgers university. AI Magazine, 6(3):192–202, 1985.

[106] Alex S. C. Lee, James H. Cutts, Gordon C. Sharp, and Joyce A. Mitchell. Ai/learn network. Journal of MedicalSystems, 11(5):349–358, October 1987.

[107] James F. Porter, C. Lawrence, KingslandIii, Donald A. B. Lindberg, Indravadan Shah, James M. Benge, Susan E.Hazelwood, Donald R. Kay, Mitsuo Homma, Masashi Akizuki, Makoto Takano, and Gordon C. Sharp. The ai/rheumknowledge-based computer consultant system in rheumatology. performance in the diagnosis of 59 connective tissuedisease patients from japan. Arthritis & Rheumatism, 31(2):219–226, 1988.

[108] S. Chokhani. Correspondences between biomathematical and causal models for clinical decision making. Journal ofMedical Systems, 5(4):249–264, December 1981.

[109] G. Banks. Artificial intelligence in medical diagnosis: the internist/caduceus approach. Critical reviews in medicalinformatics, 1(1):23–54, 1986.

[110] Randolph Miller. Computer-assisted diagnostic decision support: history, challenges, and possible paths forward.Advances in Health Sciences Education, 14(0):89–106, September 2009.

[111] G. Werner. Methuselah–an expert system for diagnosis in geriatric psychiatry. Computers and biomedical research,an international journal, 20(5):477–488, October 1987.

[112] A. S. Rigby. Development of a scoring system to assist in the diagnosis of rheumatoid arthritis. Methods ofinformation in medicine, 30(1):23–29, 1991.

[113] W. J. Long, S. Naimi, and M. G. Criscitiello. Development of a knowledge base for diagnostic reasoning in cardiology.Computers and biomedical research, an international journal, 25(3):292–311, June 1992.

[114] J. Pelz, V. Arendt, and J. Kunze. Computer assisted diagnosis of malformation syndromes: an evaluation of threedatabases (lddb, possum, and syndroc). American journal of medical genetics, 63(1):257–267, May 1996.

[115] L. A. Zadeh. Biological applications of the theory of fuzzy sets and systems. In International Symposium onBiocybernetics of Central Nervous System, pages 199–206. Little, Brown and Company, 1969.

[116] E. Sanchez. Medical Diagnosis and Composite Fuzzy Relations, pages 437–444. Amsterdam: North-Holland, 1979.

[117] L. A. Zadeh. Toward a theory of fuzzy systems, pages 209–245. Holt, Rinehart and Winston, New York, 1971.

[118] H. Leitich, H. P. Kiener, G. Kolarz, C. Schuh, W. Graninger, and K. P. Adlassnig. A prospective evaluation of themedical consultation system cadiag-ii/rheuma in a rheumatological outpatient clinic. Methods Inf Med, 40(3):213–220, Jul 2001.

[119] K. P. Adlassnig, W. Scheithauer, and G. Grabner. Computer-assisted diagnosis and its application in pancreaticdiseases. Acta Med Austriaca, 11(3-4):125–134, 1984.

[120] Agata Ciabattoni, Thomas Vetterlein, and Klaus-Peter P. Adlassnig. A formal logical framework for cadiag-2.Studies in health technology and informatics, 150:648–652, 2009.

[121] C. A. Holzmann, C. A. Perez, and E. Rosselot. A fuzzy model for medical diagnosis. Medical progress throughtechnology, 13(4):171–178, 1988.

[122] C. K. Lim, K. M. Yew, K. H. Ng, and B. J. Abdullah. A proposed hierarchical fuzzy inference system for thediagnosis of arthritic diseases. Australasian physical & engineering sciences in medicine, 25(3):144–150, September2002.

[123] Kavishwar Wagholikar, Sanjeev Mangrulkar, Ashok Deshpande, and Vijayraghavan Sundararajan. Evaluation offuzzy relation method for medical decision support. Journal of Medical Systems, pages 1–7, April 2010.

20

[124] Mohammad-R R. Akbarzadeh-T and Majid Moshtagh-Khorasani. A hierarchical fuzzy rule-based approach toaphasia diagnosis. Journal of biomedical informatics, 40(5):465–475, October 2007.

[125] M. G. Tsipouras, T. P. Exarchos, D. I. Fotiadis, A. P. Kotsia, K. V. Vakalis, K. K. Naka, and L. K. Michalis.Automated diagnosis of coronary artery disease based on data mining and fuzzy modeling. Information Technologyin Biomedicine, IEEE Transactions on, 12(4):447–458, July 2008.

[126] Satoru Sakaguchi, Katsunari Takifuji, Seizaburo Arita, and Hiroki Yamaue. Development of an early diagnosticsystem using fuzzy theory for postoperative infections in patients with gastric cancer. Digestive surgery, 21(3):210–214, 2004.

[127] T. Ichimura, E. Tazaki, and K. Yoshida. Extraction of fuzzy rules using neural networks with structure leveladaptation–verification to the diagnosis of hepatobiliary disorders. International journal of bio-medical computing,40(2):139–146, October 1995.

[128] Okure U. Obot and Faith-Michael M. Uzoka. Experimental study of fuzzy-rule based management of tropicaldiseases: case of malaria diagnosis. Studies in health technology and informatics, 137:328–339, 2008.

[129] L. Kuncheva. An aggregation of pro and con evidence for medical decision support systems. Computers in biologyand medicine, 23(6):417–424, November 1993.

[130] J. M. Barreto and F. M. de Azevedo. Connectionist expert systems as medical decision aid. Artificial intelligencein medicine, 5(6):515–523, December 1993.

[131] Elif Derya D. Ubeyli and Inan Guler. Automatic detection of erthemato-squamous diseases using adaptive neuro-fuzzy inference systems. Computers in biology and medicine, 35(5):421–433, June 2005.

[132] Charles O. Akinyokun, Okure U. Obot, Faith-Michael M. Uzoka, and John J. Andy. A neuro-fuzzy decision supportsystem for the diagnosis of heart failure. Studies in health technology and informatics, 156:231–244, 2010.

[133] Mesut Tez and Selda Tez. Neurofuzzy systems for the prediction of outcome in acute lower gastrointestinal hemor-rhage. European journal of gastroenterology & hepatology, 20(8), August 2008.

[134] L. Godo, R. L. de Mantaras, J. Puyol-Gruart, and C. Sierra. Renoir, pneumon-ia and terap-ia: three medicalapplications based on fuzzy logic. Artificial intelligence in medicine, 21(1-3):153–162, 2001.

[135] E. Binaghi, O. De Giorgi, G. Maggi, T. Motta, and A. Rampini. Computer-assisted diagnosis of postmenopausal os-teoporosis using a fuzzy expert system shell. Computers and biomedical research, an international journal, 26(6):498–516, December 1993.

[136] S. Zahan. A fuzzy approach to computer-assisted myocardial ischemia diagnosis. Artificial intelligence in medicine,21(1-3):271–275, 2001.

[137] M. D. Innis. Clinical problem solving–the role of expert laboratory systems. Medical informatics = Medecine etinformatique, 22(3):251–261, 1997.

[138] M. B. Causer, G. A. Findlay, C. R. Hawes, and D. R. Boswell. Assessment of a computerized system for the diagnosisof iron deficiency. Pathology, 26(1):37–39, January 1994.

[139] Doraid Dalalah and Sami Magableh. A remote fuzzy multicriteria diagnosis of sore throat. Telemedicine journaland e-health : the official journal of the American Telemedicine Association, 14(7):656–665, September 2008.

[140] C. A. Pena-Reyes and M. Sipper. A fuzzy-genetic approach to breast cancer diagnosis. Artificial intelligence inmedicine, 17(2):131–155, October 1999.

[141] N. Belacel, P. Vincke, J. M. Scheiff, and M. R. Boulassel. Acute leukemia diagnosis aid using multicriteria fuzzyassignment methodology. Computer methods and programs in biomedicine, 64(2):145–151, February 2001.

[142] R. Jain and A. Abraham. A comparative study of fuzzy classification methods on breast cancer data. Australasianphysical & engineering sciences in medicine / supported by the Australasian College of Physical Scientists in Medicineand the Australasian Association of Physical Sciences in Medicine, 27(4):213–218, December 2004.

[143] L. I. Kuncheva. Evaluation of computerized medical diagnostic decisions via fuzzy sets. International journal ofbio-medical computing, 28(1-2):91–100, 1991.

[144] Voula C. Georgopoulos and Chrysotomos D. Stylios. Diagnosis support using fuzzy cognitive maps combined withgenetic algorithms. In Annual International Conference of the IEEE Engineering in Medicine and Biology Society,pages 6226–6229, 2009.

21

[145] R. I. John and P. R. Innocent. Modeling uncertainty in clinical diagnosis using fuzzy logic. Systems, Man, andCybernetics, Part B: Cybernetics, IEEE Transactions on, 35(6):1340–1350, November 2005.

[146] Pasi Luukka and Tapio Leppalampi. Similarity classifier with generalized mean applied to medical data. Computersin biology and medicine, 36(9):1026–1040, September 2006.

[147] Vahid Khatibi and Gholam l. i. Montazer. Intuitionistic fuzzy set vs. fuzzy set application in medical patternrecognition. Artificial Intelligence in Medicine, 47(1):43–52, September 2009.

[148] Han-Ying Y. Kao. Diagnostic reasoning and medical decision-making with fuzzy influence diagrams. Computermethods and programs in biomedicine, 90(1):9–16, April 2008.

[149] Metin Akay, Maurice Cohen, and Donna Hudson. Fuzzy sets in life sciences. Fuzzy Sets Syst., 90(2):219–224, 1997.

[150] W. G. Baxt. Use of an artificial neural network for the diagnosis of myocardial infarction. Annals of internalmedicine, 115(11):843–848, December 1991.

[151] P. J. Lisboa. A review of evidence of health benefit from artificial neural networks in medical intervention. Neuralnetworks : the official journal of the International Neural Network Society, 15(1):11–39, January 2002.

[152] R. L. Kennedy, R. F. Harrison, A. M. Burton, H. S. Fraser, W. G. Hamer, D. MacArthur, R. McAllum, and D. J.Steedman. An artificial neural network system for diagnosis of acute myocardial infarction (ami) in the accidentand emergency department: evaluation and comparison with serum myoglobin measurements. Computer methodsand programs in biomedicine, 52(2):93–103, February 1997.

[153] M. L. Astion, M. H. Wener, R. G. Thomas, G. G. Hunder, and D. A. Bloch. Application of neural networks to theclassification of giant cell arteritis. Arthritis and rheumatism, 37(5):760–770, May 1994.

[154] J. Li, Y. Mu, and L. Zhang. [study of the pulmonary heart disease computer-aided diagnosis system based oncombining neural network]. Sheng wu yi xue gong cheng xue za zhi = Journal of biomedical engineering = Shengwuyixue gongchengxue zazhi, 18(4):573–576, December 2001.

[155] Resul Das, Ibrahim Turkoglu, and Abdulkadir Sengur. Diagnosis of valvular heart disease through neural networksensembles. Computer methods and programs in biomedicine, 93(2):185–191, February 2009.

[156] Lutz Leistritz, Miroslaw Galicki, Eberhard Kochs, Ernst Bernhard B. Zwick, Clemens Fitzek, Jurgen R. Reichenbach,and Herbert Witte. Application of generalized dynamic neural networks to biomedical data. IEEE transactions onbio-medical engineering, 53(11):2289–2299, November 2006.

[157] I. G. Vlachonikolis, D. A. Karras, M. J. Hatzakis, and N. Paritsis. Improved statistical classification methods incomputerized psychiatric diagnosis. Medical decision making : an international journal of the Society for MedicalDecision Making, 20(1):95–103, 2000.

[158] N. H. Mann and M. D. Brown. Artificial intelligence in the diagnosis of low back pain. The Orthopedic clinics ofNorth America, 22(2):303–314, April 1991.

[159] Malek Adjouadi, Melvin Ayala, Mercedes Cabrerizo, Nuannuan Zong, Gabriel Lizarraga, and Mark Rossman. Clas-sification of leukemia blood samples using neural networks. Annals of biomedical engineering, 38(4):1473–1482, April2010.

[160] Nuannuan Zong, Malek Adjouadi, and Melvin Ayala. Optimizing the classification of acute lymphoblastic leukemiaand acute myeloid leukemia samples using artificial neural networks. Biomedical sciences instrumentation, 42:261–266, 2006.

[161] Jonathan L. Jesneck, Loren W. Nolte, Jay A. Baker, Carey E. Floyd, and Joseph Y. Lo. Optimized approach todecision fusion of heterogeneous data for breast cancer diagnosis. Medical physics, 33(8):2945–2954, August 2006.

[162] Rohit Dua, Daryl G. Beetner, William V. Stoecker, and Donald C. Wunsch. Detection of basal cell carcinoma usingelectrical impedance and neural networks. IEEE transactions on bio-medical engineering, 51(1):66–71, January 2004.

[163] Angelo Andriulli, Enzo Grossi, Massimo Buscema, Alberto Pilotto, Virginia Festa, and Francesco Perri. Artifi-cial neural networks can classify uninvestigated patients with dyspepsia. European journal of gastroenterology &hepatology, 19(12):1055–1058, December 2007.

[164] Bruno Annibale and Edith Lahner. Assessing the severity of atrophic gastritis. European journal of gastroenterology& hepatology, 19(12):1059–1063, December 2007.

[165] J. S. Shang, Y. S. Lin, and A. M. Goetz. Diagnosis of mrsa with neural networks and logistic regression approach.Health care management science, 3(4):287–297, September 2000.

22

[166] E. Pesonen. Is neural network better than statistical methods in diagnosis of acute appendicitis? Studies in healthtechnology and informatics, 43 Pt A:377–381, 1997.

[167] G. Rovetta, G. Bianchi, P. Monteforte, L. Buffrini, and G. Ghirardo. Automated diagnosis and characterization oflyme disease using neural network analysis. The Journal of rheumatology, 22(3):571–572, March 1995.

[168] Enzo Grossi, Massimo P. Buscema, David Snowdon, and Piero Antuono. Neuropathological findings processed byartificial neural networks (anns) can perfectly distinguish alzheimer’s patients from controls in the nun study. BMCneurology, 7:15+, June 2007.

[169] Orhan Er, Cengiz Sertkaya, Feyzullah Temurtas, and A. Cetin Tanrikulu. A comparative study on chronic obstructivepulmonary and pneumonia diseases diagnosis using neural networks and artificial immune system. Journal of medicalsystems, 33(6):485–492, December 2009.

[170] C. Lagor, D. Aronsky, M. Fiszman, and P. J. Haug. Automatic identification of patients eligible for a pneumoniaguideline: comparing the diagnostic accuracy of two decision support models. Studies in health technology andinformatics, 84(Pt 1):493–497, 2001.

[171] Yogender Aggarwal, Bhuwan Mohan M. Karan, Barda Nand N. Das, and Rakesh Kumar K. Sinha. An unsupervisedneural network to predict the level of heat stress. Journal of clinical monitoring and computing, 22(6):425–430,December 2008.

[172] D. J. Sargent. Comparison of artificial neural networks with other statistical approaches: results from medical datasets. Cancer, 91(8 Suppl):1636–1642, April 2001.

[173] Theodore Anagnostou, Mesut Remzi, Michael Lykourinas, and Bob Djavan. Artificial neural networks for decision-making in urologic oncology. European urology, 43(6):596–603, June 2003.

[174] G. Schwarzer, W. Vach, and M. Schumacher. On the misuses of artificial neural networks for prognostic anddiagnostic classification in oncology. Statistics in medicine, 19(4):541–561, February 2000.

[175] Rudy Setiono, Wee K. Leow, and James Y. L. Thong. Opening the neural network black box: an algorithm forextracting rules from function approximating artificial neural networks. In ICIS ’00: Proceedings of the twentyfirst international conference on Information systems, pages 176–186, Atlanta, GA, USA, 2000. Association forInformation Systems.

[176] Elizabeth S. Burnside. Bayesian networks: computer-assisted diagnosis support in radiology. Academic radiology,12(4):422–430, April 2005.

[177] W. W. Chapman and P. J. Haug. Bayesian modeling for linking causally related observations in chest x-ray reports.pages 587–591, 1998.

[178] G. J. Price, W. G. McCluggage, M. L. Morrison M, G. McClean, L. Venkatraman, J. Diamond, H. Bharucha,R. Montironi, P. H. Bartels, D. Thompson, and P. W. Hamilton. Computerized diagnostic decision support systemfor the classification of preinvasive cervical squamous lesions. Human pathology, 34(11):1193–1203, November 2003.

[179] Susan M. Maskery, Hai Hu, Jeffrey Hooke, Craig D. Shriver, and Michael N. Liebman. A bayesian derived networkof breast pathology co-occurrence. Journal of biomedical informatics, 41(2):242–250, April 2008.

[180] S. Raza, Yachna Sharma, Qaiser Chaudry, Andrew N. Young, and May D. Wang. Automated classification of renalcell carcinoma subtypes using scale invariant feature transform. 2009:6687–6690, 2009.

[181] Vibha Anand and Stephen M. Downs. Probabilistic asthma case finding: a noisy or reformulation. In Annual AMIASymposium, pages 6–10, 2008.

[182] Bilal A. Ahmed, Michael E. Matheny, Phillip L. Rice, John R. Clarke, and Omolola I. Ogunyemi. A comparisonof methods for assessing penetrating trauma on retrospective multi-center data. Journal of biomedical informatics,42(2):308–316, April 2009.

[183] D. Aronsky, M. Fiszman, W. W. Chapman, and P. J. Haug. Combining decision support methodologies to diagnosepneumonia. pages 12–16, 2001.

[184] In Sook S. Cho and Peter J. Haug. The contribution of nursing data to the development of a predictive model forthe detection of acute pancreatitis. Studies in health technology and informatics, 122:139–142, 2006.

[185] W. J. Long, H. Fraser, and S. Naimi. Reasoning requirements for diagnosis of heart disease. Artificial intelligencein medicine, 10(1):5–24, May 1997.

[186] M. Korver and P. J. Lucas. Converting a rule-based expert system into a belief network. Medical informatics =Medecine et informatique, 18(3):219–241, 1993.

23

[187] M. Suojanen, S. Andreassen, and K. G. Olesen. A method for diagnosing multiple diseases in munin. IEEEtransactions on bio-medical engineering, 48(5):522–532, May 2001.

[188] David Foreman, Stephanie Morton, and Tamsin Ford. Exploring the clinical utility of the development and well-being assessment (dawba) in the detection of hyperkinetic disorders and associated diagnoses in clinical practice.Journal of child psychology and psychiatry, and allied disciplines, 50(4):460–470, April 2009.

[189] M. A. Shwe, B. Middleton, D. E. Heckerman, M. Henrion, E. J. Horvitz, H. P. Lehmann, and G. F. Cooper.Probabilistic diagnosis using a reformulation of the internist-1/qmr knowledge base. i. the probabilistic model andinference algorithms. Methods of information in medicine, 30(4):241–255, October 1991.

[190] Bastian Wemmenhove, Joris Mooij, Wim Wiegerinck, Martijn Leisink, Hilbert Kappen, and Jan Neijt. Inferencein the promedas medical expert system. In Artificial Intelligence in Medicine, Lecture Notes in Computer Science,chapter 61, pages 456–460. 2007.

[191] M. Cleret, F. Le Duff, A. Fresnel, and P. Le Beux. Diamed: a probabilistic diagnostic aid system on the web. Studiesin health technology and informatics, 84(Pt 1):429–433, 2001.

[192] Zuoshuang Xiang, Rebecca M. Minter, Xiaoming Bi, Peter J. Woolf, and Yongqun He. minituba: medical inferenceby network integration of temporal data using bayesian analysis. Bioinformatics (Oxford, England), 23(18):2423–2432, September 2007.

[193] S. Wong, D. Wu, and Y. Yao. Critical remarks on the computational complexity in probabilistic inference. InGuoyin Wang, Qing Liu, Yiyu Yao, and Andrzej Skowron, editors, Rough Sets, Fuzzy Sets, Data Mining, andGranular Computing, volume 2639 of Lecture Notes in Computer Science, chapter 114, pages 587–588–588. SpringerBerlin / Heidelberg, Berlin, Heidelberg, April 2003.

[194] Kwokleung Chan, Te-Won W. Lee, Pamela A. Sample, Michael H. Goldbaum, Robert N. Weinreb, and Terrence J.Sejnowski. Comparison of machine learning and traditional classifiers in glaucoma diagnosis. IEEE transactions onbio-medical engineering, 49(9):963–974, September 2002.

[195] Joarder Kamruzzaman and Rezaul K. Begg. Support vector machines and other pattern recognition approaches tothe diagnosis of cerebral palsy gait. IEEE transactions on bio-medical engineering, 53(12 Pt 1):2479–2490, December2006.

[196] B. R. Brewer, S. Pradhan, G. Carvell, and A. Delitto. Feature selection for classification based on fine motor signsof parkinson’s disease. 2009:214–217, 2009.

[197] C. Lu, T. Van Gestel, J. A. Suykens, S. Van Huffel, I. Vergote, and D. Timmerman. Preoperative predictionof malignancy of ovarian tumors using least squares support vector machines. Artificial intelligence in medicine,28(3):281–306, July 2003.

[198] Francesco Camastra and Alessandro Verri. A novel kernel method for clustering. IEEE transactions on patternanalysis and machine intelligence, 27(5):801–805, May 2005.

[199] Elisabetta La Torre, Tatiana Tommasi, Barbara Caputo, and Giovanni E. Gigante. Kernel methods for melanomarecognition. Studies in health technology and informatics, 124:983–988, 2006.

[200] Qianfei Yuan, Congzhong Cai, Hanguang Xiao, Xinghua Liu, and Yufeng Wen. Svm-aided cancer diagnosis basedon the concentration of the macroelement and microelement in human blood. Sheng wu yi xue gong cheng xue zazhi = Journal of biomedical engineering = Shengwu yixue gongchengxue zazhi, 24(3):513–518, June 2007.

[201] Jennifer Listgarten, Sambasivarao Damaraju, Brett Poulin, Lillian Cook, Jennifer Dufour, Adrian Driga, JohnMackey, David Wishart, Russ Greiner, and Brent Zanke. Predictive models for breast cancer susceptibility frommultiple single nucleotide polymorphisms. Clinical cancer research : an official journal of the American Associationfor Cancer Research, 10(8):2725–2737, April 2004.

[202] Baek Hwan H. Cho, Hwanjo Yu, Kwang-Won W. Kim, Tae Hyun H. Kim, In Young Y. Kim, and Sun I. Kim.Application of irregular and unbalanced data to predict diabetic nephropathy using visualization and feature selectionmethods. Artificial intelligence in medicine, 42(1):37–53, January 2008.

[203] K. Kabasawa and S. Kaihara. A sequential diagnostic model for medical questioning. Medical informatics =Medecine et informatique, 6(3):175–185, 1981.

[204] B. S. Duran and T. O. Lewis. An application of cluster analysis to the construction of a diagnostic classification.Computers in biology and medicine, 4(2):183–188, December 1974.

[205] R. Thurmayr, M. Otte, and R. Thurmayr. [computer-aided diagnosis for pancreatic function test (author’s transl)].Langenbecks Archiv fur Chirurgie, 339:253–257, November 1975.

24

[206] D. P. McKenzie, P. D. McGorry, C. S. Wallace, L. H. Low, D. L. Copolov, and B. S. Singh. Constructing a minimaldiagnostic decision tree. Methods of information in medicine, 32(2):161–166, April 1993.

[207] B. Sahiner, H. P. Chan, N. Petrick, R. F. Wagner, and L. Hadjiiski. Feature selection and classifier performance incomputer-aided diagnosis: the effect of finite sample size. Medical physics, 27(7):1509–1522, July 2000.

[208] W. J. Long, J. L. Griffith, H. P. Selker, and R. B. D’Agostino. A comparison of logistic regression to decision-treeinduction in a medical domain. Computers and biomedical research, an international journal, 26(1):74–97, February1993.

[209] M. Drent, M. A. van Nierop, F. A. Gerritsen, E. F. Wouters, and P. G. Mulder. A computer program using balf-analysis results as a diagnostic tool in interstitial lung diseases. American journal of respiratory and critical caremedicine, 153(2):736–741, February 1996.

[210] C. L. Tsien, H. S. Fraser, W. J. Long, and R. L. Kennedy. Using classification tree and logistic regression methodsto diagnose myocardial infarction. Studies in health technology and informatics, 52 Pt 1:493–497, 1998.

[211] S. M. Rudolfer, G. Paliouras, and I. S. Peers. A comparison of logistic regression to decision tree induction in thediagnosis of carpal tunnel syndrome. Computers and biomedical research, an international journal, 32(5):391–414,October 1999.

[212] S. Kable, R. Henry, R. Sanson-Fisher, M. Ireland, R. Corkrey, and J. Cockburn. Childhood asthma: can computersaid detection in general practice? The British journal of general practice : the journal of the Royal College ofGeneral Practitioners, 51(463):112–116, February 2001.

[213] E. Wayne Holden, Elizabeth Grossman, Hoang Thanh T. Nguyen, Margaret J. Gunter, Becky Grebosky, AnnVon Worley, Leila Nelson, Scott Robinson, and David J. Thurman. Developing a computer algorithm to identifyepilepsy cases in managed care organizations. Disease management : DM, 8(1):1–14, February 2005.

[214] R. J. Marshall. Partitioning methods for classification and decision making in medicine. Statistics in medicine,5(5):517–526, 1986.

[215] K. J. Cios, R. E. Freasier, L. S. Goodenday, and L. T. Andrews. An expert system for diagnosis of coronary arterystenosis based on 201tl scintigrams using the dempster-shafer theory of evidence. Computer applications in thebiosciences : CABIOS, 6(4):333–342, October 1990.

[216] N. Matsuda. Computer-assisted laboratory diagnosis. Rinsho byori. The Japanese journal of clinical pathology,39(3):243–251, March 1991.

[217] T. Sekiya, A. Watanabe, and M. Saito. The use of modified constellation graph method for computer-aided classi-fication of congenital heart diseases. IEEE transactions on bio-medical engineering, 38(8):814–820, August 1991.

[218] J. C. Toha, S. Vasquez, P. Fuentes, and M. A. Soto. Algorithm for assisting medical diagnosis. Computer methodsand programs in biomedicine, 39(3-4):303–309, April 1993.

[219] R. Stamper, B. S. Todd, and P. Macpherson. Case-based explanation for medical diagnostic programs, with anexample from gynaecology. Methods of information in medicine, 33(2):205–213, May 1994.

[220] C. D. Evans and R. M. Winter. A case-based learning approach to grouping cases with multiple malformations.M.D. computing : computers in medical practice, 12(2):127–136, 1995.

[221] S. Dzeroski and N. Lavrac. Rule induction and instance-based learning applied in medical diagnosis. Technologyand health care, 4(2):203–221, August 1996.

[222] M. C. Jaulent, C. Le Bozec, E. Zapletal, and P. Degoulet. Case based diagnosis in histopathology of breast tumours.Studies in health technology and informatics, 52 Pt 1:544–548, 1998.

[223] A. S. Ochi-Okorie. Disease diagnosis validation in tropix using cbr. Artificial intelligence in medicine, 12(1):43–60,January 1998.

[224] E. Armengol, A. Palaudaries, and E. Plaza. Individual prognosis of diabetes long-term risks: a cbr approach.Methods of information in medicine, 40(1):46–51, March 2001.

[225] G. I. Paterson. A rough sets approach to patient classification in medical records. Medinfo. MEDINFO, 8 Pt 2,1995.

[226] W. Z. Liu, A. P. White, M. T. Hallissey, and J. W. Fielding. Machine learning techniques in early screening forgastric and oesophageal cancer. Artificial intelligence in medicine, 8(4):327–341, August 1996.

25

[227] K. Viikki, E. Kentala, M. Juhola, and I. Pyykko. Decision tree induction in the diagnosis of otoneurological diseases.Medical informatics and the Internet in medicine, 24(4):277–289, 1999.

[228] M. Zorman, H. P. Eich, P. Kokol, and C. Ohmann. Comparison of three databases with a decision tree approach inthe medical field of acute appendicitis. Studies in health technology and informatics, 84(Pt 2):1414–1418, 2001.

[229] M. R. Ortız-Posadas, J. F. Martınez-Trinidad, and J. Ruız-Shulcloper. A new approach to differential diagnosis ofdiseases. International journal of bio-medical computing, 40(3):179–185, January 1996.

[230] P. Ramnarayan, N. Cronje, R. Brown, R. Negus, B. Coode, P. Moss, T. Hassan, W. Hamer, and J. Britto. Vali-dation of a diagnostic reminder system in emergency medicine: a multi-centre study. Emerg Med J, 24(9):619–624,September 2007.

[231] Kaizhu Huang, Haiqin Yang, Irwin King, and Michael R. Lyu. Maximizing sensitivity in medical diagnosis usingbiased minimax probability machine. IEEE transactions on bio-medical engineering, 53(5):821–831, May 2006.

[232] Albert M. Lai, Simon Parsons, and George Hripcsak. Fuzzy temporal constraint networks for clinical information.pages 374–378, 2008.

[233] Guenter Tusch, Chris Bretl, Martin O’Connor, Martin Connor, and Amar Das. SPOT–towards temporal datamining in medicine and bioinformatics. 2008.

[234] P. Ruch and Section Editor for the IMIA Yearbook Section on Decision Support. A Medical Informatics Perspectiveon Decision Support. Toward a Unified Research Paradigm Combining Biological vs. Clinical, Empirical vs. Legacy,and Structured vs. Unstructured Data. Yearbook of medical informatics, pages 96–98, 2009.

[235] R. A. Miller, M. A. McNeil, S. M. Challinor, F. E. Masarie, and J. D. Myers. The internist-1/quick medical referenceproject–status report. The Western journal of medicine, 145(6):816–822, December 1986.

[236] D. Aronsky and P. J. Haug. Diagnosing community-acquired pneumonia with a bayesian network. In AMIASymposium, pages 632–636, Dept. of Medical Informatics, LDS Hospital/University of Utah, Salt Lake City, USA.,1998.

[237] Kavishwar B. Wagholikar and Ashok W. Deshpande. Fuzzy relation based modeling for medical diagnostic decisionsupport: Case studies. International Journal of Knowledge-based and Intelligent Engineering Systems, 12(5,6):319–326, 2008.

[238] Kavishwar B. Wagholikar, Sundararajan Vijayraghavan, and Ashok W. Deshpande. Fuzzy naive bayesian model formedical diagnostic decision support. In 31st Annual International Conference of the IEEE Engineering in Medicineand Biology Society, volume 1, pages 3409–3412, September 2009.

[239] G. H. Guyatt, R. B. Haynes, R. Z. Jaeschke, D. J. Cook, L. Green, C. D. Naylor, M. C. Wilson, andW. S. Richardson.Users’ guides to the medical literature: Xxv. evidence-based medicine: principles for applying the users’ guides topatient care. evidence-based medicine working group. JAMA : the journal of the American Medical Association,284(10):1290–1296, September 2000.

[240] J. E. Wennberg. Dealing with medical practice variations: a proposal for action. Health affairs (Project Hope),3(2):6–32, 1984.

[241] J. C. Horrocks, G. Devroede, and F. T. de Dombal. Computer-aided diagnosis of gastroenterologic diseases insherbrooke: preliminary report. Canadian journal of surgery. Journal canadien de chirurgie, 19(2):160–164, March1976.

[242] P. Haug, P. D. Clayton, P. Shelton, T. Rich, I. Tocino, P. R. Frederick, R. O. Crapo, W. J. Morrison, and H. R.Warner. Revision of diagnostic logic using a clinical database. Medical decision making : an international journalof the Society for Medical Decision Making, 9(2):84–90, 1989.

[243] Eric J. Horvitz, John S. Breese, and Max Henrion. Decision theory in expert systems and artificial intelligence. Int.J. Approx. Reasoning, 2(3):247–302, July 1988.

[244] E. H. Shortliffe, S. G. Axline, B. G. Buchanan, T. C. Merigan, and S. N. Cohen. An artificial intelligence programto advise physicians regarding antimicrobial therapy. Comput Biomed Res, 6(6):544–560, December 1973.

[245] David Heckerman. Probabilistic interpretations for mycin’s certainty factors. In Proceedings of the 1st AnnualConference on Uncertainty in Artificial Intelligence (UAI-85), New York, NY, 1985. Elsevier Science.

[246] Jeffrey A. Barnett. Computational methods for a mathematical theory of evidence. In IJCAI, pages 868–875.William Kaufmann, 1981.

26

[247] Benjamin N. Grosof. Evidential confirmation as transformed probability. In John F. Lemmer and Laveen Kanal,editors, First International Workshop on Uncertainty in Artificial Intelligence, volume 1, Amsterdam, Netherlands,August 1986. North Holland (Elsevier Science).

[248] D. E. Heckerman and R. A. Miller. Towards a better understanding of the internist-1 knowledge base. In Medinfo86,pages 27–31, New-york, 1986. North Holland.

[249] Amos Tversky and Daniel Kahneman. Judgement under uncertainty: Heuristics and biases. Science, 185(4157):1124–1131, September 1974.

[250] David Heckerman. An empirical comparison of three inference methods. In Proceedings of the 4th Annual Conferenceon Uncertainty in Artificial Intelligence (UAI-88), New York, NY, 1988. Elsevier Science.

[251] J. L. Liu, J. C. Wyatt, J. J. Deeks, S. Clamp, J. Keen, P. Verde, C. Ohmann, J. Wellwood, M. Dawes, and D. G.Altman. Systematic reviews of clinical decision tools for acute abdominal pain. Health Technol Assess, 10(47),November 2006.

[252] E. S. Berner, G. D. Webster, A. A. Shugerman, J. R. Jackson, J. Algina, A. L. Baker, E. V. Ball, C. G. Cobbs, V. W.Dennis, and E. P. Frenkel. Performance of four computer-based diagnostic systems. N Engl J Med, 330(25):1792–1796, June 1994.

[253] D. Aronsky, K. J. Chan, and P. J. Haug. Evaluation of a computerized diagnostic decision support system forpatients with pneumonia: study design considerations. Journal of the American Medical Informatics Association :JAMIA, 8(5):473–485, 2001.

[254] Laura J. van ’t Veer, Hongyue Dai, Marc J. van de Vijver, Yudong D. He, Augustinus A. Hart, Mao Mao, Hans L.Peterse, Karin van der Kooy, Matthew J. Marton, Anke T. Witteveen, George J. Schreiber, Ron M. Kerkhoven,Chris Roberts, Peter S. Linsley, Rene Bernards, and Stephen H. Friend. Gene expression profiling predicts clinicaloutcome of breast cancer. Nature, 415(6871):530–536, January 2002.

[255] Jose Palma, Jose M. Juarez, Manuel Campos, and Roque Marin. Fuzzy theory approach for temporal model-baseddiagnosis: An application to medical domains. Artificial intelligence in medicine, 38(2):197–218, October 2006.

[256] D. Kopecky, M. Hayde, A. R. Prusa, and K. P. Adlassnig. Knowledge-based interpretation of toxoplasmosis serologytest results including fuzzy temporal concepts–the toxonet system. Medinfo, 10(Pt 1):484–488, 2001.

[257] O. Bouhaddou, G. E. Morgan, J. G. Lambert, and D. Sorenson. Qualitative analysis of temporal information iniliad: implications for linking iliad to an electronic medical record as a knowledge server. pages 199–203, 1996.

[258] Carlo Combi, Elpida K. Papailiou, and Yuval Shahar. Temporal Information Systems in Medicine. Springer Pub-lishing Company, Incorporated, 2010.

[259] Juan C. Augusto. Temporal reasoning for decision support in medicine. Artif. Intell. Med., 33(1):1–24, 2005.

[260] K. P. Adlassnig, C. Combi, A. K. Das, E. T. Keravnou, and G. Pozzi. Temporal representation and reasoning inmedicine: Research directions and challenges. Artif Intell Med, 38(2):101–113, October 2006.

[261] George J. Klir and Bo Yuan. Fuzzy Sets and Fuzzy Logic: Theory and Applications. Prentice Hall PTR, 1st edition,May 1995.

[262] T. Fawcett. An introduction to roc analysis. Pattern Recognition Letters, 27(8):861–874, June 2006.

[263] David J. Hand and Robert J. Till. A simple generalisation of the area under the roc curve for multiple classclassification problems. Mach. Learn., 45(2):171–186, November 2001.

[264] F. T. De Dombal. The diagnosis of acute abdominal pain with computer assistance: worldwide perspective. Annalesde chirurgie, 45(4):273–277, 1991.

[265] F. T. De Dombal, Susan E. Clamp, Angela Softley, Biba J. Unwin, and John R. Staniland. Prediction of individualpatient prognosis: Value of computer-aided systems. Med Decis Making, 6(1):18–22, February 1986.

[266] Davide Luciani, Silvio Cavuto, Luca Antiga, Massimo Miniati, Simona Monti, Massimo Pistolesi, and GuidoBertolini. Bayes pulmonary embolism assisted diagnosis: a new expert system for clinical use. Emergency medicinejournal : EMJ, 24(3):157–164, March 2007.

[267] Rosa Blanco, Inaki Inza, Marisa Merino, Jorge Quiroga, and Pedro Larranaga. Feature selection in bayesian classifiersfor the prognosis of survival of cirrhotic patients treated with tips. Journal of biomedical informatics, 38(5):376–388,October 2005.

27

[268] Sholom M. Weiss and Casimir A. Kulikowski. Expert: a system for developing consultation models. In IJCAI’79:Proceedings of the 6th international joint conference on Artificial intelligence, pages 942–947, San Francisco, CA,USA, 1979. Morgan Kaufmann Publishers Inc.

[269] Tang-Kai K. Yin and Nan-Tsing T. Chiu. A computer-aided diagnosis for distinguishing tourette’s syndrome fromchronic tic disorder in children by a fuzzy system with a two-step minimization approach. IEEE transactions onbio-medical engineering, 51(7):1286–1295, July 2004.

[270] P. H. Bartels, D. Thompson, and J. E. Weber. Diagnostic decision support by inference networks. In vivo (Athens,Greece), 7(4):379–385, 1993.

[271] C. E. Kahn, L. M. Roberts, K. Wang, D. Jenks, and P. Haddawy. Preliminary investigation of a bayesian networkfor mammographic diagnosis of breast cancer. In Annual Symposium on Computer Application in Medical Care,pages 208–212, 1995.

[272] R. E. Bolinger, K. J. Hopfensperger, and D. F. Preston. Application of a virtual neurode in a model thyroiddiagnostic network. pages 310–314, 1991.

[273] E. L. Kinney, R. J. Wright, and J. W. Caldwell. A classifier system for the diagnosis of disease from routinelaboratory values. Journal of medical systems, 12(5):319–326, October 1988.

[274] Moshe Ben-Bassat, Richard W. Carlson, Venod K. Puri, Mark D. Davenport, John A. Schriver, Mohamed Latif,Ronald Smith, Larry D. Portigal, Edward H. Lipnick, and Max H. Weil. Pattern-based interactive diagnosis ofmultiple disorders: The medas system. Pattern Analysis and Machine Intelligence, IEEE Transactions on, PAMI-2(2):148–160, 1980.

[275] K. Smith, A. Clark, K. Dyson, E. Kruger, L. Lejmanoski, A. Russell, and M. Tennant. Guided self diagnosis: aninnovative approach to triage for emergency dental care. Australian dental journal, 51(1):11–15, March 2006.

[276] Norman E. Betaque and G. Anthony Gorry. Automating judgmental decision making for a serious medical problem.Management Science, 17(8):B–421–434, April 1971.

[277] P. W. Hamilton, R. Montironi, W. Abmayr, M. Bibbo, N. Anderson, D. Thompson, and P. H. Bartels. Clinicalapplications of bayesian belief networks in pathology. Pathologica, 87(3):237–245, June 1995.

[278] C. E. Kahn. Artificial intelligence in radiology: decision support systems. Radiographics : a review publication ofthe Radiological Society of North America, Inc, 14(4):849–861, July 1994.

[279] Guillaume Marrelec, Philippe Ciuciu, Melanie Pelegrini-Issac, and Habib Benali. Estimation of the hemodynamicresponse in event-related functional mri: Bayesian networks as a framework for efficient bayesian modeling andinference. IEEE transactions on medical imaging, 23(8):959–967, August 2004.

[280] F. T. De Dombal, J. C. Horrocks, J. R. Staniland, and P. J. Guillou. Production of artificial ”case histories” byusing a small computer. British medical journal, 2(5761):578–581, June 1971.

[281] Vipin Kumar, Ananth Grama, Anshul Gupta, and George Karpis. Introduction to Parallel Computing: Design andAnalysis of Parallel Algorithms. Benjamin-Cummings Pub Co, January 1994.

[282] Karen Sandler, Lysandra Ohrstrom, Laura Moy, and Robert McVay. Killed by code: Software transparency inimplantable medical devices — opinion — communications of the acm, July 2010.

[283] Michael Swash and Michael Glynn. Hutchison’s Clinical Methods: An Integrated Approach to Clinical Practice WithSTUDENT CONSULT Online Access. Saunders Ltd., April 2007.

[284] D. Hunscher, A. Boyd, L. A. Green, and D. J. Clauw. Representing natural-language case report form terminologyusing health level 7 common document architecture, loinc, and snomed-ct: lessons learned. AMIA Annu Symp Proc,2006.

[285] C. Friedman. A broad-coverage natural language processing system. pages 270–274, 2000.

[286] Elizabeth S. Chen, George Hripcsak, and Carol Friedman. Disseminating natural language processed clinical narra-tives. pages 126–130, 2006.

[287] S. Goryachev, M. Sordo, and Q. T. Zeng. A suite of natural language processing tools developed for the i2b2 project.AMIA Annu Symp Proc, 2006.

[288] Olivier Bodenreider. The unified medical language system (umls): integrating biomedical terminology. Nucl. AcidsRes., 32(suppl 1):D267–270, January 2004.

28