MEDICAL KNOWLEDGE MANAGEMENT

MEDICAL KNOWLEDGE MANAGEMENT BY USING DATA MINING METHOD: THE CASE OF PARKINSON’S DISEASE
WU, SHIANG-HAU
Faculty of Management and Administration, Macau University of Science and Technology, Avenida Wai Long, Taipa, Macau SAR, China
E-mail:shwu@must.edu.mo
WEN, YA-YUN
Department of Agribusiness Management, National Pingtung University of Science and Technology, 1, Shuefu Road, Neipu
Pingtung, 912, Taiwan, China
E-mail: edwinawen@hotmail.com
Clinical decision- making needs available information to be the guidance for physicians. Nowadays, data mining method is applied in medical research in order to analyze large volume of medical data. This study attempts to use data mining method to analyze the databank of Parkinson’s disease and to explore whether the voice measurement variables can be the diagnostic tool for the Parkinson’s disease.
1. Introduction
1.1. Medical Knowledge Management and Data Mining
In clinical research, medical information is essential for diagnosis and patient care. For clinical research, it also provides useful information to facilitate therapeutic improvement and conduct medical researches. The medical knowledge management in the realm of medical information can be shown as the cycle among the clinical research, guidelines, quality indicators, performance measures, outcomes and the concept (McCourt et. al. 2007). In order to integrate clinical information management, medical data analysis, and application development, clinical decision intelligence (CDI) is emerged as the new area to streamline the data management from clinical practice, nursing, health-care management, health-care administration. As for the CDI, knowledge discovery is used in the knowledge acquisition and the evidence-based research stage to analyze the information extracted from research reports, reports, evidence tables, flow charts, guidelines that include evidence contents, sources and quality scores (Wang et. al., 2007).
1.2. The Parkinson’s Disease Case
The Parkinson’s disease (PD) is a type of neurological disease. Many neurological diseases affect phonation of patients, and voice can be a valuable aid in the diagnosis of neurological disease (Michelsson et. al., 1982, Ramig et. al., 1988). In Pakinson’s disease, voice disorders affect approximately 45% of patients (Logemann et. al., 1978). Previous studies have shown that PD is associated with vowel prolongation, syllable repetitions, isolated sentences and conversation. Rosen et.al. (2005) explores that syllable repetitions (diadochokinesis, DDK) is particularly useful for describing intensity decay of language ability associated with PD. Other recent studies are related to the voice treatment of PD (Spielman et.al., 2007; Baudelle et. al., 2003). Although previous studies offer some useful information for PD diagnosis, whether voice measurements can be the suitable tool for diagnosis needs to be examined. This study aims at examining whether by solely using vocal measurements can researchers discriminate PD patients from healthy people.
2. Method
The study applies several analysis methods, including factor analysis, logistic regression, decision tree and neural net, to analyze the dataset of PD. The goals of this study include the following aspects, (i) Examine the biomedical voice measurements by three data mining methods and find out which voice measurement (and component) would significantly discriminate PD patients from healthy people. (ii) Examine the application of these data mining methods to the PD dataset and find out which methods would have the lowest Type 1 and Type 2 errors.
2.1 Data
This dataset, offered by Max Little of the University of Oxford, in collaboration with the National Centre for Voice and Speech, Denver, Colorado, who recorded the speech signals, is composed of a range of biomedical voice measurements from 31 people, 23 with Parkinson's disease (PD). Each column in the table is a particular voice measure, and each row corresponds one of 195 voice recording from these individuals ("name" column). The main aim of the data is to discriminate healthy people from those with PD, according to "status" column which is set to 0 for healthy and 1 for PD. The original study published the feature extraction methods for general voice disorders (Little et al., 2008).
The variables in this dataset are listed as follows,
• Name--subject name and recording number
• MDVP:Fo(Hz)- Average Vocal Fundamental Frequency (V1)
• MDVP:Fhi(Hz)- Maximun Vocal Fundamental Frequency (V2)
• MDVP: Flo(Hz)- Minimun Vocal Fundamental Frequency (V3)
• MDVP:Jitter(%), MDVP:Jitter(Abs), MDVP:RAP, MDVP:PPQ, Jitter:DDP- several measures of variation in fundamental frequency (V4, V5, V6, V7, V8)
• MDVP:Shimmer, MDVP:Shimmer(db), Shimmer:APQ3, Shimmer:APQ5, MDVP:APQ, Shimmer:DDA- several measures of variation in amplitude (V9, V10, V11, V12, V13, V14)
• NHR, HNR- Two measures of ratio of noise to tonal components in voice
• Status- Health status of the subject, Parkinson’s (one), Healthy (zero)
• RPDE,D2- Two nonlinear dynamical complexity measures
• DFA – Signal fractal scaling component
• Spread 1, spread 2, PPE- Three nonlinear measures of fundamental frequency variation

2.2 Validity
The concept of validity is composed of internal validity and external validity. Internal validity is the extent to which a study provides evidence of a cause-effect relationship between the independent and dependent variables. Campbell and Stanley (1966) indicate threats of internal validity take place when researchers can not randomly assign participants to treatments.
External validity denotes whether causal relationships can be generalized into different measures, persons, settings and times. As for the public health research, researchers tend to be more concerned with the internal validity than the external validity. It results in the failure to translate research into public health practice (Steckler and McLeroy, 2008). External validity is often neglected by researchers, funding agencies, ethics committees, the pharmaceutical industry, medical journals and governmental regulators, leaving clinicians make judgments (Rothwell, 2005). Steckler and McLeroy (2008) quote four categories of external validity which should be reported on journal papers,
(i) Study participant recruitment and selection procedures, participant rates, and representative nature at the levels of individuals, intervention staff, and delivery settings.
(ii) Level and consistency of implementation across program components, settings, staff and time.
(iii) Impact on a variety of outcomes, especially those important to populations, practitioners, and decision makers.
(iv) Follow-up reports should include all levels in item (i), long term effects on outcomes in item (iii), and program sustainability, modification, or discontinuance.
The two types of validity of this study are discussed as follows,
(i) Internal validity: According to Little et. al.(2008), the databank was originated from the study of sustained phonation tests due to the concept that simple sustained phonation tests are able to elicit dysphonic symptoms. Although whether the participants were randomly assigned in depends on the research design of Little et. al.(2008), the cause-effect relationship between independent and dependent variables is presented in the following analysis in this study.
(ii) External validity: External validity of the study relates to the research process in Little et. al. (2008). According to Little et. al.(2008), the databank consists of 195 sustained vowel phonations from 31 male and female subjects, of which 23 were diagnosed of PD. The time since PD patients were diagnosed ranged from 0 to 28 years, and ages of the subjects ranged from 46 to 85 years (mean 65.8, standard deviation 9.8). Although the related information of external validity is insufficient in Little et. al.(2008) and the generalization of causal relationship in this study is restrained due to the sample size and its focus on the middle age and aged population, this study offers useful information for clinical reference.
2.3 Factor Analysis
The goal of factor analysis is to find out the characteristics of the variables in the databank. Factor analysis in multivariate techniques is used commonly. Factor analysis is a mathematical tool which can be used to examine large data set and utilize the entire correlation among variables to find the communalities (Sharma, 1996).
There are three most popular methods in factor analysis: principal component factoring (PCF), principal axis factoring (PAF) and maximum likelihood estimation (MLE). The PCF method is introduced as follows,
Consider a p-variable and q-factor model given by the following equations,
Z1= a11b1+ a12b2+….+ a1qbq+ε1
Z2= a21b1+ a22b2+….+ a2qbq+ε2 (1)
.
.
Zp= ap1b1+ ap2b2+….+ apqbq+εp
Where Z1…Zp are p variables (i.e. variables in the databank), apq is the pattern loading of the p-th variable on the q-th factor, andεp is the unique factor for the p-th variable. Eq(1) can be written as :
Z’= cb’+ε’ (2)
Where Z’ is a p ×1 vector of variables, c is the p × q matrix of factor loadings, b’ is a q×1 vector of unobserved factors, and ε’is a p ×1 vector of unique factors.
2.3.1 Factor Analysis Results
The study uses SPSS 10.0 software to analyze 22 voice measurement variables (except status variable), and gets the following results:
(i) KMO Bartlett Test: KMO =0.886>0.80. It means there are communalities among variables, and it is suitable to conduct the factor analysis.
(ii) Communalities: The result of communalities is listed below in Table 1.
Table 1: Communalities of voice measurement variables
Initial Extraction
V1 1.000 .849
V2 1.000 .499
V3 1.000 .596
V4 1.000 .984
V5 1.000 .964
V6 1.000 .984
V7 1.000 .950
V8 1.000 .984
V9 1.000 .969
V10 1.000 .967
V11 1.000 .927
V12 1.000 .968
V13 1.000 .920
V14 1.000 .927
NHR 1.000 .895
HNR 1.000 .814
RPDE 1.000 .632
DFA 1.000 .717
SPREAD1 1.000 .839
SPREAD2 1.000 .598
D2 1.000 .649
PPE 1.000 .821
Table 1 shows the communalities of the voice measurement variables is larger than 0.20. These variables can be retained.
(iii) The result of principal component analysis: In the principal component analysis, the common method is the eigen-value-greater-than-one rule and the scree plot. According to the result of principal component analysis and the following scree plot, four components are retained. The four components can explain 83.868% of the total variance.

















Fig.1: Scree plot of voice measurement variables
As for the principal component analysis, the orthogonal rotation and varimax methods are widely used. The objective of rotation is to achieve a simpler factor structure that can be meaningfully interpreted by the researcher. An orthogonal rotation can be performed to achieve this objective. In the orthogonal rotation, varimax and quartimax are most popular types, which the rotated factors are orthogonal to each other (Jollife, 2002). The results of orthogonal rotation and varimax methods are presented in Table 2.
Table 2: The Factor Loadings after Varimax Rotating
Component
1 2 3 4
V12 .883 .407 .109 -.103
V9 .853 .466 .141 -6.741E-02
V13 .837 .435 .168 -3.746E-02
V11 .832 .464 .113 -7.707E-02
V14 .832 .464 .113 -7.708E-02
V10 .831 .512 .114 -4.642E-02
HNR -.712 -.437 -.313 -.131
D2 .584 .113 .291 .459
V8 .397 .900 .112 6.270E-02
V6 .398 .900 .112 6.270E-02
V4 .409 .887 .167 4.099E-02
V5 .279 .873 .327 -.128
V7 .483 .835 .134 -3.464E-02
NHR .386 .822 .125 .233
V3 8.462E-02 -8.000E-02 -.753 .122
RPDE .311 .145 .714 6.761E-02
SPREAD1 .442 .452 .657 -8.951E-02
V1 .158 -.140 -.642 .626
SPREAD2 .466 5.177E-02 .613 4.635E-02
PPE .494 .474 .578 -.135
DFA .154 5.435E-02 -1.714E-02 -.831
V2 -3.008E-02 .141 -.126 .680

In Table 2, the component 1 includes V9, V10, V11, V12, V13, V14, HNR, D2. Because the component 1 is mainly composed of variables about variation in amplitude, the component 1 can be renamed as the variation in amplitude. The component 2 includes V4, V5, V6, V7, V8, NHR. Because the component 2 is mainly composed of variables about several measures of variation in fundamental frequency, the component 2 can be renamed as the variation in fundamental frequency. The component 3 includes SPREAD1, SPREAD2, V1, V3, RPDE, PPE. Because the component 3 is mainly composed of three nonlinear measures of fundamental frequency variation, and V1, V3 are the average and the minimum vocal fundamental frequency, the component 3 can be renamed as the nonlinear measures of fundamental frequency variation. The component 4 includes DFA, V2. Because two variables have larger difference, the component 4 can be renamed as the other measure of voices.
2.3.2 Logistic Regression Results
The study attempts to examine whether physicians can finely diagnose PD solely by means of the voice measurements. Therefore, the study uses the logistic regression to examine the odds of correct diagnosis of PD. The study uses status variable as the dependent variable and the component 1 to the component 4 resulted from the factor analysis as the covariates to construct the logistic regression model. The results are listed below,
(i) Cox & Snell R- square and Nagelkerke R-square: In the logistic regression model, the Cox & Snell R- square is 0.350 and the Nagelkerke R-square is 0.521. It means these four components of voice measurement variables have strong relationship with the health status variables.
(ii) Hosmer-Lemeshow test: Hosmer and Lemeshow (2000) stipulate the Hosmer -Lemeshow test in order to examine whether the logistic model is well fitted. If the p-value of Hosmer-Lemeshow test is larger than 0.05, it means the model is well fitted. In this logistic model, Chi-square value is 6.605, p-value is 0.580>0.05. So the logistic regression model is well fitted.
(iii) Classification: The logistic regression model offers the prediction of classification. The classification result is shown in Table 3:
Table 3: Classification Result

Predicted
STATUS Percentage Correct
Observed .0000000 1.0000000
Step 1 STATUS .0000000 26 22 54.2
1.0000000 11 136 92.5
Overall Percentage 83.1
In Table 2, 162 (=26+136) cases are correctly classified into healthy groups and PD patients group, while healthy cases are falsely classified into PD patients group, and 11 PD cases are falsely classified into the healthy group. The correct percentage of classification is 83.1%. Besides, Type 1 error probability (healthy people falsely classified as PD patients) is 45.8 %and Type 2 error probability (PD patients falsely classified as healthy people) is 7.5%.
(iv) The result of the logistic regression: The coefficients evaluation is presented in Table 4:
Table 4: The Logistic Regression Result

B S.E. Wald df Sig. Exp(B)
component1 1.817 .472 14.823 1 .000 6.152
component2 .424 .436 .942 1 .332 1.527
component 3 1.479 .265 31.099 1 .000 4.389
component4 -.515 .228 5.114 1 .024 .597
Constant 2.117 .365 33.557 1 .000 8.308

In Table 4, the significance ratio of the Wald Test in component 2 (Variation in Fundamental Frequency) is larger than 0.05. Other components, including component 1, component 3, component 4, are significant (p<0.05). So component 1(variation in amplitude), component 3 (nonlinear measures of fundamental frequency variation), and component 4(other measure of voices) are important variables to predict and explain the healthy status.
According to the results of the logistic regression, the odd ratio can be calculated from Table 4. The odd ratio of component 1 is 6.152. It means when the component 1 increases 0.01 units, the probability of the odd between PD cases and healthy cases increases 0.01× (6.152-1)= 5.152%. In the same way, when the component 3 increases 0.01 units, the probability of the odd between PD cases and healthy cases increases 0.01× (4.389-1) =3.389%. Besides, when the component 4 increases 0.01 units, the probability of the odd between PD cases and healthy cases increases 0.01× (0.597-1) =-0.00403%.
2.4 Decision Tree Analysis
Decision tree analysis is useful for logical induction in the data mining process. Decision tree induction is the learning of decision trees from class-labeled training tuples. A decision tree is a flowchart- like tree structure, where each internal node represents a test on an attribute, each branch represents an outcome of the test, and each leaf node (terminal node) holds a class label. The topmost node is the root node (Han and Kamber, 2006).
Rattle 2.4.78 software is applied to the decision tree analysis in this study. The healthy status variable is the response variable, and the 22 voice measurement variables are the input variables in the decision tree model. In the decision tree analysis, 70% of samples (136 cases) are applied and the remaining 30% of the samples are used as test cases for evaluation. After considering the suitability for explanation, the conditional inference tree model is used here. The configuration parameters is originally set by the Rattle 2.4.78 software, including min. split=20, max. Depth=30 and min. bucket=7. The decision tree is shown in the Fig.2.
In Fig.2, the first node follows two decisions. When the variable SPREAD 1 is less than or equal to -6.324, 43 cases are classified into the healthy group (status=0). If the variable SPREAD is larger than -6.324, other unclassified cases would fall into the third node. The third node follows two decisions. If the RPDE is less than or equal to 0.398, 8 cases are classified into the special group, while these members have 50% probability belong to the PD group or the healthy group. If the variable RPDE is larger than 0.398, 85 cases are classified into the PD group.

Fig.2: Decision Tree of the Parkinson’sDisease Case

The overall error probability of classification for the trained cases is 14.70%, and 8.47% for the test cases. The error matrix of the trained cases and the test cases are listed in Table 5 and Table 6.
Table 5: Error matrix for the trained cases in decision tree model (counts)
Actual
STATUS
Predicted .0000000 1.0000000
STATUS 0.3720930 27 16
0.5 4 4
0.9764705 2 83
Table 6: Error matrix for the test cases in decision tree model (counts)
Actual
STATUS
Predicted .0000000 1.0000000
STATUS 0.3720930 10 4
0.5 1 3
0.9764705 4 37

As for the calculation of Type 1 and Type 2 errors, the predicted status=0.3720930 is considered as status=0, and the predicted status=0.9764705 is considered as status=1. The predicted status=0.5 in both trained and test cases are regarded as falsely classified due to the uncertain explanation for the results. So the Type 1 error probability is (4+2+1+4)/195=5.64%, and Type 2 error probability is (16+4+4+3)/195=13.85%.
2.5 Neural Net Analysis
A neural network is a set of connected input and output units in which each connection has a weight associated with it. During the learning phase, the network learns by adjusting the weights so as to be able to predict the correct class label of the input tuples. Neural network learning is also referred to as connectionist learning due to the connections between units (Han and Kamber, 2006). The goal of the neural net analysis is to build a model that is based on the idea of multiple layers of neurons connected to each other, feeding the numeric data through the network, combining the numbers, to produce a final answer.
In the neural net model, 70% of the samples (136 cases) are applied and the remaining 30% of the samples are used as test cases for evaluation. According to the results calculated by Rattle 2.4.78 software, the training pattern of the neural net would be a 216-1-1 network (216 units, 1 hidden layer node and 1 output level) with 435 weights if considering the square measures of areas under the ROC (Receiver Operating Characteristic) curves among possible neural net models. If the square measure approaches to 0.5, it would be the less corresponding model. If the square measure equals to 1, it would bet the model with perfect accuracy. According to the calculation of Rattle 2.4.78 software, the square measure of the area under the ROC curve is 0.5886, which is the largest among all possible neural net models. The ROC curve of the neural net model in the study is shown in Fig.3.

Fig.3: ROC Curve of the Neural Net Model

The overall error probability of classification for the trained cases is 9.56%, and 25.42% for the test cases. The error matrix of the neural net model is shown in Table 7 and Table 8

Table 7: Error matrix for the trained cases of the Neural Net model (counts)
Actual
STATUS
Predicted .0000000 1.0000000
STATUS .0000000 20 0
1.0000000 13 103

Table 8: Error matrix for the test cases of the Neural Net model (counts)

Actual
STATUS
Predicted .0000000 1.0000000
STATUS .0000000 3 3
1.0000000 12 41

From Table 7 and Table 8,Type 1 error probability is (13+12)/195=12.82%, and Type 2 error probability is 3/195=1.53%.
3. Discussion
The study applies the factor analysis, the logistic regression method, the decision tree model and the neural net model to analyze whether voice measurement variables can discriminate PD patients from healthy people. The major results are listed below,
(i) According to the results of the factor analysis and the logistic regression model, the component 2 (Variation in Fundamental Frequency) is insignificant. It represents that jitter, the traditional measurement method evaluating the extent of variation, doesn’t discriminate PD patients significantly in the PD case. According to Little et.al. (2008), noise-to-harmonics ratios variables (NHR, HNR) are also belong to traditional measurement variables. In the factor analysis result, NHR is one of the elements of the component 2, which is insignificant in the logistic regression model. Therefore, NHR is also insignificant. The result meets the result of Little et.al. (2008) in SVM classification performance results.
(ii) The result of the logistic regression model also indicates the component 1(variation in amplitude) and the component 3 (nonlinear measures of the fundamental frequency variation) have the positive relationship with the odd probability between the healthy group and the PD patients group.
(iii) Little et.al.(2008) indicates that vocal production is a highly nonlinear dynamical system, and that changes caused by impairments to the vocal organs, muscles and nerves will affect the dynamics of the whole system. In the nonlinear measurement variables, SPREAD1 and RPDE are two important nodes in the decision tree model. In the decision tree model, the value of SPREAD 1 is the criterion to classify the healthy people in the whole sample. The value of RPDE can also be the criterion to classify the PD cases from the other members in the sample. The result partially meets the conclusion of Little et. al. (2008), which estimates that PPE produces the best performance in classification and the combination of HNR, RPDE, DFA and PPE obtains best overall classification performance.
(iv) According to the results of all three methods, Type 1 and Type 2 errors are listed in Table 9:



Table 9: Type 1 and Type 2 errors in three methods
Type 1 error Type 2 error
Logistic Regression 45.8% 7.5%
Decision Tree Model 5.64% 13.85%
Neural Net Model 12.82% 1.53%

According to Table 9, the logistic regression model has the highest Type 1 error and medical professionals can’t solely depend on the logistic regression method for PD diagnosis if considering the PD dataset. The decision tree model has the lowest Type 1 error, but has the highest Type 2 error. It means the decision tree model is not the best tool for medical professionals to discriminate PD patients from healthy people. The neural net model has the lowest Type 2 error in all three models. Since it has larger Type 1 error (12.82%), its lowest Type 2 error vindicates that neural net model has better performance compared with other traditional statistical methods in classification (Boritz, 1995).
4. Conclusion
The study uses the data mining analysis in knowledge discovery to explore the Parkinson’s Disease data. The study finds out some new information about the performance of different data mining methods compared with the original study of this dataset. Nowadays, data mining is widely used in the realm of the preventive medicine. In further study, medical researchers can develop the evaluation table according to the results of data mining in order to make physicians and ordinary people aware the early symptoms of PD and make earlier treatments. Although the sample size and the experimental process of the databank restrain the generalization of the study, the study offers useful information for clinical reference regarding the PD diagnosis for the middle age and the aged people.

Acknowledgments
The authors are grateful to two anonymous reviewers for the comments on this paper.
References
Baudelle, E., Vassiere, J., Renard, J.L., Roubeau, B., Chevrie-Mueller, C. (2003)” Carateristiques Vocaliques Intrinseques et Co-intinseques dans les dysarthries cerebelleuses et parkinsonienne”, Folia Phoniatrica et Logopaedica, 55, 137-146.
Boritz, J.E. (1995), “Effectiveness of Neural Network Types for Prediction of Business Failure”, Expert Systems with Applications, 9, 503-512.
Campbell, D.T., Stanley, J.C. (1966), Experimental and Quasi Experimental Designs, Chicago: Rand McNally,.
Han, J., Kamber, M. (2006), Data Mining: Concepts and Techniques, CA: Elsevier.
Hosmer, D.W. , Lemeshow, S.(2000). Applied Logistic Regression, Second Edition, New York: John Wiley & Sons.
Jollife, I.T.(2002), Principal Component Analysis, Second Edition, New York: Springer-Verlag.
Little, M.A., McSharry, P.E., Hunter, E.J., Ramig, L.O. (2008), “Suitability of Dysphonia Measurements for Telemonitoring of Parkinson’s Disease”, IEEE Transactions on Biomedical Engineering (to appear).
Logemann, J., Fisher, H., Boshes, B., Blonsky, R.E.(1978)”Frequency and Concurrence of Vocal Tract Dysfunctions in the Speech of a Large Sample of Parkinson Patients”, Journal of Speech Hear Disord , 43, 47-57.
McCourt, B., Harrington, R.A., Fox, K., Hamilton, C.D., Booher, K., Hammond, W.E., Walden, A., Nahm, M. (2007) “Data Standards: At the Intersection of Sites, Clinical Research Networks, and Standards Development Initiatives”, Drug Information Journal, 41(3):393–404.
Michelsson, K., Raes, J., Thoden, C., Wasz-Hockert, O. (1982)”Sound Spectrographic Cry Analysis in Neonatal Diagonostics: An Evaluative Study”, Journal of Phonet, 10, 79-88.
Ramig, L., Sherer, R., Titze, I., Ringel, S. (1988)”Acoustic Analysis of Voices of Patients with Neurologic Disease: Rationale and Preliminary Data”, Ann Otol Rhinol Larygnol, 97, 164-172.
Rosen, K.M., Kent, R.D., Duffy, J.R. (2005)” Task-Based Profile of Vocal Intensity Decline in Parkinson’s Disease”, Folia Phoniatrica et Logopaedica, 57, 28-37.
Rothwell, P.M. (2005)”Treating Individuals 1: External Validity of Randomized Controlled Trials: To whom do the results of this trials apply?”, The Lancet, 365(9453), 82-93.
Sharma, S. (1996) Applied Multivariate Techniques, John Wiley & Sons, Inc.
Spielman, J., Ramig, L.O., Maeler, L., Halpern, A., William, J. (2007)”Effects of an Extended Version of the Lee Silverman Voice Treatment on Voice and Speech in Parkinson’s Disease”, Language Pathology, 16(2), 95-107.
Steckler, A. and McLeroy, K.R. (2008), “The Importance of External Validity”, American Journal of Public Health, 98(1), 9-10.
Wang, X.S., Nayda, L. and Dettinger, R. (2007) “Infrastructure for a Clinical Decision-Intelligence System”, IBM Systems Journal, 46(1), 151-169.