INTELLIGENT CLINICAL DECISION SUPPORT SYSTEMS

Alexandru G. Floares

SAIA - Solutions of Artiﬁcial Intelligence Applications, Str. Vlahuta, Bloc Lama C/45, Cluj-Napoca, Romania

IOCN - Cancer Institute Cluj-Napoca, Artiﬁcial Intelligence Department, Str. Republicii, Nr. 34-36, Cluj-Napoca, Romania

Keywords:

Clinical decision support systems, Data mining, Artiﬁcial intelligence, Chronic hepatitis, Prostate cancer,

Biopsy.

Abstract:

Clinical Decision Support Systems (CDSS) have the potential to replace painful, invasive, and costly proce-

dures, to optimize medical decisions, improve medical care, and reduce costs. An even better strategy is to

make use of a knowledge discovery in data approach, with the aid of artiﬁcial intelligence tools. This results

in transforming conventional CDSS in Intelligent Clinical Decision Support (i-CDSS). Evolving i-CDSS give

to the conventional CDSS the capability of self-modifying their rules set, through supervised learning from

patients data. Intelligent and evolving CDSS represent a strong foundation for evidence-based medicine. We

proposed a methodology of building i-CDSS and related concepts. These are illustrated with some of our

results in liver diseases and prostate cancer, some of them showing the best published performance.

1 INTRODUCTION AND

BACKGROUND

The use of information technology for replacing

painful, invasive, and/or costly procedures, for opti-

mizing various medical decisions, or for improving

medical care and reducing the costs, represent major

goals of Medical Informatics. Implementing Clini-

cal Decision Support Systems (CDSS) (Berner, 2007)

on a large scale is a major step toward these goals.

Using a knowledge discovery in data approach with

artiﬁcial intelligence tools, one can build Intelligent

Clinical Decision Support Systems (i-CDSS) instead

of conventional CDSS.

We performed a set of investigations on con-

structing i-CDSS for several liver diseases, prostate

and thyroid cancer, and chromosomal disorders (e.g.,

Down syndrome) during pregnancy. The high perfor-

mance of the liver and prostate i-CDSS determined

us to consider some of these methods and concepts

mature and general enough to be presented, but still

under development. Here, we present a methodology

of building i-CDSS and the related concepts. These

are illustrated with some of our results in liver and

prostate diseases, showing the best published perfor-

mances out to date, to our knowledge.

Chronic Hepatitis B and C are major diseases of

mankind and a serious global public health problem.

The persons with these chronic diseases are at high

risk of death from cirrhosis and liver cancer. Liver

biopsy is the gold standard for grading the severity

of disease, and staging the degree of ﬁbrosis, and the

grade of necroinﬂammation. The most used scoring

systems are:

1. METAVIR A (A stands for activity) or Ishak NI

(NI stands for necroinﬂammatory) for necroin-

ﬂammatory grade

2. METAVIR F or Ishak F for the ﬁbrosis stage (F

stands for ﬁbrosis).

By assigning scores for severity, grading, and staging

of hepatitis, they are very important for patient man-

agement.

Liver biopsy is invasive, painful, and relatively

costly; complications severe enough to require hos-

pitalization can occur in about 4% of patients (Lin-

dor, 1996). In a review of over 68,000 patients re-

covering from liver biopsy, 96% experienced adverse

symptoms during the ﬁrst 24 hours of recovery. Hem-

orrhage was the most common symptom, but infec-

tions also occurred. Side effects of the biopsies in-

cluded pain, tenderness, internal bleeding, pneumoth-

orax, and rarely, death (Tobkes and Nord, 1995)

There are two main non-invasive diagnosis tech-

niques of interest (Shaheen et al., 2007). FibroScan is

282

G. Floares A. (2010).

INTELLIGENT CLINICAL DECISION SUPPORT SYSTEMS.

In Proceedings of the Third International Conference on Health Informatics, pages 282-287

DOI: 10.5220/0002740802820287

 SciTePress

a type of ultrasound machine that uses transient elas-

tography to measure liver stiffness. The device re-

ports a value that is measured in kilopascals (kPa). Fi-

broTest for assessing ﬁbrosis, and ActiTest for assess-

ing necroinﬂammatory activity are available through

BioPredictive (www.biopredictive.com). These tests

use algorithms to combine the results of serum tests

of beta 2-macroglobulin, haptoglobulin, apolipopro-

tien A1, total bilirubin, gamma glutamyltranspepti-

dase (GGT), and alanine aminotransferase (ALT).

The results of these diagnosis techniques are not

directly interpretable by a pathologist, but can be ex-

trapolated to a ﬁbrosis and necroinﬂammation score .

FibroTest and FibroScan have reasonably good utility

for the identiﬁcation of cirrhosis, but lesser accuracy

for earlier stages. It is considered that reﬁnements are

necessary before these tests can replace liver biopsy

(Shaheen et al., 2007).

We used a knowledge discovery in data, based on

artiﬁcial intelligence, to investigatethe possibilities of

accuracy improvements and of expressing the results

in the pathologist scoring systems.

In our prostate cancer studies, one goal was to in-

vestigate the possibility of developing a non-invasive

diagnosis i-CDSS, based mainly on the concentra-

tions of a set of 8 angiogenic molecules in serum.

A detailed description of these data can be found in

(Balacescu et al., 2008). For the purpose of this pa-

per, which is to outline the methodology and the con-

cepts related to evolving intelligent CDSS, unnec-

essary molecular biology, medical and data mining

technicalities were eliminated.

To our knowledge, this is the ﬁrst study, inte-

grating angiogenic molecules, clinical and laboratory

data, to develop intelligent systems capable to predict

diagnosis like prostate cancer or benignant diseases,

which are usually based on the prostate biopsy. The

preliminary results are very encouraging, with an ac-

curacy ranging from 97.8% to 100% (manuscript in

preparation).

2 DEVELOPING INTELLIGENT

CLINICAL DECISION

SUPPORT: A METHODOLOGY

AND RELATED CONCEPTS

In essence, we extracted and integrated information

from various non-invasive data sources, e.g. imag-

ing, clinical, routine laboratory or molecular data, and

build i-CDSS capable to predict various results of the

biopsy, e.g., liver ﬁbrosis or prostate cancer diagnosis,

with an acceptable accuracy. The meaning of accept-

able accuracy depends on the speciﬁc medical context

and is a matter of consensus. Probably, in this context

the prediction accuracy should be at least 80%.

Because the results of the liver, prostate, or other

organs biopsy, are used in many important medical

decisions, in the management of the related diseases,

we investigated the possibility of developing other i-

CDSS, starting from these. For example, important

treatment decision are partially based on biopsy.

Chronic hepatitis B and C are treated with drugs

called Interferon or Lamivudine, which can help some

patients. This treatment decision is based on several

patients selection criteria. For example, the criteria

for selecting the patients with chronic hepatitis C who

will beneﬁt from Interferon treatment, are:

1. Chronic infection with hepatitis C virus (HCV):

antibodies against HCV (anti-HCV) are present

for at least 3 months.

(a) the hepatitis B surface antigen (HBsAg) is

present for at least 6 months, or

(b) the hepatitis B e antigen (HBeAg) is present for

at least 10 weeks.

2. The cytolytic syndrome: the transaminases level

is increased or normal.

3. Pathology (biopsy): the Ishak NI ≥ 4 and Ishak F

≤ 3.

4. The virus is replicating: the transaminases level

is increased or normal, and anti-HCV are present,

and RNA-HCV ≥ 10

copies/milliliter.

For hepatitis B there is a similar set of selection

rules.

Analyzing these treatment decisions, one can

identify two problems:

1. Invasiveness: the patients selection criteria in-

clude ﬁbrosis and necroinﬂammation assessed by

liver biopsy, an invasive medical procedure.

2. Cost of the wrong decisions: patients selec-

tion errors are very costly, because Interferon or

Lamivudine therapy costs thousands of dollars.

In the methodological context proposed in this

paper, developing solutions to these problems is

straightforward. The aforementioned conditions are

easy to implement in an interactivecomputer program

and biopsy could be replaced by the non-invasive i-

Biopsy. Developing intelligent CDSS, based on non-

invasive medical investigations, and optimized selec-

tion criteria, could be of great beneﬁt to the patients

and could also save money.

More precisely, one should investigate if it is pos-

sible:

INTELLIGENT CLINICAL DECISION SUPPORT SYSTEMS

283

1. To build i-CDSS capable to predict the biopsy

results—ﬁbrosis stage and necroinﬂammation

grade—with an accuracy of at least 80%.

2. To integrate the i-CDSS predicting the biopsy re-

sults with the other selection criteria in an i-CDSS

for Interferon treatment.

3. To make the Interferon treatment i-CDSS an

evolving one, capable to optimize the treatment

decisions by self-modifying through learning.

It is interesting to note that some of the compo-

nents of the treatment i-CDSS are ﬁxed in time, while

other can evolve, by learning from data. In this exam-

ple, the i-Biopsy component is ﬁxed, being an input-

output relationship, already discovered from patients

data, and used to predict the results of the biopsy,

without performing it. The component implementing

the treatment selection criteria could be either ﬁxed or

evolving through learning. In the ﬁrst case we have a

ﬁxed i-CDSS, and in the second case an evolving one.

Evolving i-CDSS can minimize the costs due to

erroneous patients selection, and maximize the ben-

eﬁt of the treatment. They can optimize the selec-

tion rule sets by ﬁnding the relevant selection criteria

and their proper cutoff values. For this, the outcomes

of the Interferon treatment must be clearly deﬁned as

numerical or categorical attributes and registered in a

data base for each treated patient.

Then, intelligent agents are employed to learn the

prediction of the treatment outcomes. They must

be capable of expressing the extracted information

as rules, using non-invasive clinical, laboratory and

imaging attributes as inputs. Using feature selection

(see for example (Guyon et al., 2006)) one will ﬁnd

the relevant patient selection criteria.

Thus, the i-CDSS started with the accepted pa-

tients selection criteria, but these are evolving. It is

worth to mention that the evolved selection criteria

could be different, from those initially proposed by

physicians, and usually better. However, they should

be always evaluated by the experts. In the supervised

learning process, intelligent agents also discover the

proper cutoff values of the relevant selection criteria.

Again, these are usually better than those proposed

by experts, but they should always be evaluated by

them. In our opinion, evolving through learning from

patients data is crucial for evidence based-medicine.

These i-CDSS are the result of a data mining pre-

dictive modeling strategy, which is now patent pend-

ing, consisting mainly in:

1. Extracting and integration information from var-

ious medical data sources, after a laborious pre-

processing:

(a) cleaning features and patients,

(b) various treating of missing data,

(d) selecting features,

(e) balancing data.

2. Testing various classiﬁers or predictive modeling

algorithms.

3. Testing various ensemble methods for combining

classiﬁers.

For modeling, we tested the prediction accuracy of

various types of artiﬁcial intelligence agents:

1. Neural Networks of various types and architec-

tures,

2. Decision trees C5.0 and Classiﬁcation and Re-

gression Trees

3. Support Vector Machines, with various kernels

4. Bayesian Networks

5. Genetic Programming based agents.

i-Biopsy is an intelligent system based on any algo-

rithm or combination of algorithms capable of learn-

ing from data. Of course, accuracy is very important

but physicians also prefer white-box algorithms and

transparent decisions.

We have chosen C5.0 decision trees, the last ver-

sion of the C4.5 algorithm (Quinlan, 1993), with

10-fold cross-validation. As ensemble method, we

used Freund and Schapire’s boosting (Freund and

Schapire, 1997) for improving the predictive power of

C5.0 classiﬁer learning systems. A set of C5.0 classi-

ﬁers is produced and combined by voting, and by ad-

justing the weights of training cases. We suggest that

boosting should always be tried when peak predictive

accuracy is required, especially when unboosted clas-

siﬁers are already quite accurate .

Genetic programming was another important

choice, giving accurate and transparent i-CDSS in the

form of mathematical models of the input-output rela-

tionship (manuscript in preparation). Transparency of

the i-CDSS is affected by boosting and is less useful

when the number of variables is large.

3 MAIN RESULTS

In what follows, some of the results illustrating the

practical and conceptual signiﬁcance of i-Biopsy are

presented. The examples are i-CDSS selected from

gastroenterology and urology.

In one of our hepatological studies, we collected a

dataset of 700 chronic hepatitis C patients and about

135 inputs. One of the i-CDSS has liver ﬁbrosis as the

HEALTHINF 2010 - International Conference on Health Informatics

284

predicted output expressed as Metavir F score, having

ﬁve classes: from Metavir F0 to Metavir F4. The ac-

curacy of the ﬁrst experiments was about 60%. Pre-

processing increased the accuracy with 20% to 25%.

As we mentioned, we tested various algorithms and

settings, but C5.0 accuracy was one of the highest,

about 80% (see Table 1 and Table 2 for some exam-

ples).

Table 1: Results of experiments for feature and algorithm

selection for METAVIR F0 prediction, formulated as a one-

versus-all classiﬁcation (Logistic Regres stands for Logistic

Regression; AUC stands for Area Under the Curve).

Model Accuracy% Features AUC

C5 99.639 19 1

CART 95.307 17 0.755

Logistic Regres 93.863 25 0.918

Neural Net 92.78 25 0.648

Table 2: Results of experiments for feature and algorithm

selection for METAVIR F4 prediction, formulated as a one-

versus-all classiﬁcation (Logistic Regres stands for Logistic

Regression; AUC stands for Area Under the Curve).

Model Accuracy% Features AUC

C5 95.668 11 0.896

CART 94.224 21 0.914

Logistic Regres 92.419 25 0.906

Neural Net 88.809 25 0.867

Parameter tuning and boosting increase the accu-

racy of some i-CDSS even to 100% (Floares et al.,

2008; Floares, 2009b).

We also developed liver i-Biopsy versions based

on Bayesian Networks and on Genetic Programming,

some of them as binary classiﬁers (work in progress).

After many experiments, we conclude that important

is to reach the highest robust accuracy, between 80%

and 100%, and the test for this is the external valida-

tion.

Because the results of the biopsy are central

to important medical decisions, in the management

of chronic hepatitis patients, it was relatively easy

to build i-CDSS for Interferon treatment (Floares,

2008), (Floares, 2009a). This was done just by adding

the aforementioned patients selection criteria to the

i-Biopsy. This non-invasive i-CDSS is of a special

kind; being able to evolve, by attempting to predict

the progressively accumulating outcomes of the Inter-

feron treatment, it will eventually identify the proper

patients selection criteria, and their cutoff values from

data (see section 2 for more details). Thus, the rules

set of this i-CDSS is evolving.

We tried to develop not only the technical aspects

of the intelligent CDSS, evolving through learning

from data, but also the related concepts.

i-Biopsy, one of the central concepts, is an intelli-

gent system (the preﬁx ”i-” coming from intelligent),

supporting important medical decisions, by being ca-

pable to predict, with an acceptable accuracy, the re-

sults usually givenby a pathologist, examining the tis-

sue samples from biopsies. Real biopsy is performed

on different organs, e.g., liver, prostate, etc., and the

pathologists expressed their ﬁndings as diagnoses or

scores of a largely accepted scoring system. While the

concept is general, individual systems must be spe-

ciﬁc (see below).

For example, let us shortly analyze liver (organ) i-

Biopsy (the intelligent counterpart of the real biopsy),

in chronic hepatitis C (disease), assessing liver ﬁbro-

sis (diagnose), expressed by METAVIR F (pathologist

scoring system). This liver i-Biopsy takes as inputs

and integrate various routine, non-invasive, clinical,

imaging and lab data.

To distinguish between the scores of the real

biopsy and their counterparts predicted by i-Biopsy,

we proposed the general terms of i-scores. There are

many examples, like Gleason score in prostate cancer,

but we continue to focus on the gastroenterological

applications, where we have:

1. The liver i-Biopsy is the i-CDSS correspond-

ing to the real liver biopsy; the i-METAVIR F

scores are the values predicted by i-Biopsy for the

METAVIR-F ﬁbrosis scores, designating exactly

the same pathological features.

2. The i-METAVIR F scores and the biopsy

METAVIR F scores could have different values

for the same patient, at the same moment, de-

pending for example on the prediction accuracy

or pathologist errors.

3. i-METAVIR F scores are obtained in a non-

invasive, painless, and riskless manner, as op-

posed to METAVIR-F scores, assessed by liver

biopsy.

For simplicity, we referred only to the METAVIR

F scores, but these considerations are general, and can

be easily extrapolated to other liver scores, like Ishak

F, METAVIR A, and Ishak NI. Moreover, these con-

ceptual clariﬁcations apply to any situation in which

the following elements are present:

1. an anatomical structure, e.g., liver, or prostate, etc.

2. the invasive procedure, e.g., biopsy

3. a disease, e.g., chronic hepatitis C or B, prostate

cancer or benign prostatic hyperplasia

4. a set of pathological features or diagnoses to as-

sess, e.g., ﬁbrosis, necroinﬂammation, etc.

INTELLIGENT CLINICAL DECISION SUPPORT SYSTEMS

285

5. a set of classes for the pathological ﬁndings, e.g.

Gleason or METAVIR scores, pathological diag-

noses, etc.

The second element (biopsy) could be replaced,

at least in a considerable percentage of cases, by the

i-Biopsy, and the ﬁfth element by the i-scores.

We have built the following i-CDSS which can be

used for Interferon treatment decision support:

1. Module for liver ﬁbrosis prediction,

(a) according to METAVIR F scoring system, with

and without liver stiffness (FibroScan),

(b) according to Ishak F scoring system, with and

without liver stiffness (FibroScan).

2. Module for the grade of necroinﬂammation (ac-

tivity) prediction, according to Ishak NI scoring

systems.

Also, we developed some prostate i-Biopsy sys-

tems, as non-invasive i-CDSS counterparts for some

prostate biopsy results. For example, i-Gleason score

is the i-Biopsy predicted Gleason score (work in

progress), and is also central to many important med-

ical decisions in prostate cancer. The three classes

classiﬁers, distinguishing between normal, benignant

and malignant are more interesting for the fundamen-

tal research. The binary classiﬁers, especially those

distinguishing between malignant and benignant, are

clinically oriented. The preliminary results are very

encouraging, with accuracy ranging from 97.8% to

100%.

4 DISCUSSION

A short digression about the meaning of the diagno-

sis accuracy, of the i-CDSS in general and i-Biopsy in

particular, seems necessary, because it confused many

physicians, especially when reporting very high val-

ues like 100%. Many physicians believe that 100%

accuracy is not possible in medicine. The mean-

ings will be made clear trough examples. Typically,

an invasive liver (or prostate, etc.) biopsy is per-

formed to the patient, and a pathologist analyzes the

tissue samples assessing ﬁbrosis, necroinﬂammation,

etc., and expressed the results as scores or patholog-

ical diagnosis. The pathologist may have access to

other patient medical data, but usually these are not

necessary for him or her to formulate the patholog-

ical diagnosis. Moreover, in some studies it is re-

quired that the pathologist knows nothing about the

patient. His or her diagnosis can be more or less cor-

rect or even wrong, for many reasons not discussed

here. We have proposed i-CDSS predicting the ﬁbro-

sis scores resulted from liver biopsy, or the prostate

cancer diagnosis resulted from prostate biopsy, with

accuracy reaching 90% - 100%. For the i-CDSS, sev-

eral clinical, imaging, and lab data of the patient are

essential, because they were somehow incorporated

in the system. They were used like input features

to train the system, and they are required for a new,

unseen patient, because i-Biopsy is a relationship be-

tween these inputs and the ﬁbrosis, necroinﬂamma-

tion scores, or diagnosis as outputs. The category of i-

CDDSs discussed here do not deal directly with diag-

nosis correctness, but with diagnosis prediction accu-

racy. Without going into details, this is due in part to

the supervised nature of the learning methods used to

build them. The intelligent agents learned to predict

the results of the biopsy given by the pathologist, and

the pathologist diagnosis could be more or less cor-

rect. For example, let us suppose that the pathologist

diagnosis is wrong. The i-Biopsy could still be 100%

accurate in predicting this wrong diagnosis, but this

is rarely the case. In other words, the i-Biopsy will

predict, in a non-invasive and painless way, and with-

out the risks of the biopsy, a diagnosis which could be

even 100% identical with the pathologist diagnosis,

if the biopsy is performed. While the accuracy and

the correctness of the diagnosis are related in a sub-

tle way, they are different matters. i-Biopsy will use

the information content of several non-invasiveinves-

tigations, to predict the pathologist diagnosis, without

performing the biopsy. The correctness of the diagno-

sis is a different matter, but typically a good accuracy

correlates well with a correct diagnoses. The accu-

racy of the diagnosis, as well as other performance

measures like the area under the receiver operating

characteristic (AUROC), for a binary classiﬁer sys-

tem (Fawcett, 2004), are useful for intelligent sys-

tems comparison. To our knowledge, the proposed

liver i-Biopsy system outperformed the most popu-

lar and accurate system, FibroTest and ActiTest (Sha-

heen et al., 2007) commercialized by BioPredictive

company, and FibroScan. The liver i-Biopsy is a

multi-classes classiﬁer, expressing the results in the

pathologist’s scoring systems, e.g., ﬁve classes for

METAVIR F and seven classes for Ishak F. Multi-

classes classiﬁers are more difﬁcult to develop than

binary classiﬁers, with outputs not directly related to

the ﬁbrosis scores. We also build binary classiﬁers as

decision trees with similar accuracy and mathemati-

cal models (work in progress). Despite the fact that

AUROC is only for binary classiﬁers, loosely speak-

ing a 100% accuracy n classes classiﬁer is equivalent

with n binary classiﬁers with AUROC = 1 (maximal).

BioPredictive company analyzed a total of 30 studies

HEALTHINF 2010 - International Conference on Health Informatics

286

(Poynard et al., 2007) which pooled 6,378 subjects

with both FibroTest and biopsy (3,501 chronic hep-

atitis C). The mean standardized AUROC was 0.85

(0.82-0.87). The robustness of these results is clearly

demonstrated by this cross-validation, while i-Biopsy

results need to be cross-validated. The fact that i-

Biopsy , in its actual setting, relies on routine ul-

trasound features is both a strong point and a weak

one, because of the subjectiveness in ultrasound im-

ages interpretation. It is worth to note that in certain

circumstances the result of the liver i-Biopsy could

be superior to that of real biopsy. When building

the i-CDSS, the results of the potentially erroneous

biopsies, which are not fulﬁlling some technical re-

quirements, were eliminated from the data set. Thus,

the i-Biopsy predicted results correspond only to the

results of the correctly performed biopsies, while a

number of of the real biopsy results are wrong, be-

cause they were not correctly performed. Due to the

invasive and unpleasant nature of the biopsy, is very

improbable that a patient will accept a technically in-

correct biopsy to be repeated. Unlike real biopsy,

i-Biopsy can be used to evaluate ﬁbrosis evolution,

which is of interest in various biomedical and pharma-

ceutical studies, because, being non-invasive,painless

and without any risk, can be repeated as many time as

needed.

Also, in the early stages of liver diseases, often the

symptoms are not really harmful for the patient, but

the treatment is more effective than in more advanced

ﬁbrosis stages. The physician will hesitate to indicate

an invasive, painful and risky liver biopsy, and the pa-

tients are not as worried about their disease as they are

about the pain of the biopsy. However, i-Biopsy can

be performed and an early start of the treatment could

be much more effective.

ACKNOWLEDGEMENTS

We thank to the following medical teams: Dr. Mon-

ica Lupsor, Dr. T. Suteu and Prof. Dr. R. Badea, from

Medical Imaging Department, Dr. H. Stefanescu and

Dr. Z. Sparchez, from Hepatology Department, Dr.

A. Serban from Pathology Department, Dr. N. Crisan,

Dr. B. Feciche and Prof. Dr. I. Coman, from Urology

Department, University of Medicine and Pharmacy

Cluj-Napoca, Romania, Dr. Carmen Floares, Dr. O.

Balcescu, Dr. Ioana Neagoe Dr. Loredana Balacescu,

Dr. Oana Tudoran, and Prof. Dr. A. Irimie, from Can-

cer Institute Cluj-Napoca, Romania. We also thank

to the computer science team: Dr. F. Manolache, E.

Suica, and T. Popa.

REFERENCES

Balacescu, O., Neagoe, I., Balacescu, L., Crisan, N., Feci-

che, B., Tudoran, O., Coman, I., and Irimie, A. (2008).

Angiogenesis serum protein quantifcation for prostate

pathology. Curr Urol, (2):181–187.

Berner, E., editor (2007). Clinical decision support systems.

Springer: New York, NY.

Fawcett, T. (2004). Roc graphs: Notes and practical con-

siderations for researchers. technical report. Technical

report, Palo Alto, USA: HP Laboratories.

Floares, A. G. (2008). Intelligent systems for interferon

treatment decision support in chronic hepatitis c based

on i-biopsy. In IDAMAP - Intelligent Data Analysis in

Biomedicine and Pharmacology, Washington DC.

Floares, A. G. (2009a). Intelligent clinical decision supports

for interferon treatment in chronic hepatitis c and b

based on i-biopsy. In International Joint Conference

on Neural Networks, 2009, Atlanta, Georgia, USA.

Floares, A. G. (2009b). Liver i-Biopsy and the Correspond-

ing Intelligent Fibrosis Scoring Systems: i-Metavir F

and i-Ishak F, pages 253–264. Lecture Notes in Com-

puter Science. Springer Berlin / Heidelberg.

Floares, A. G., Lupsor, M., Stefanescu, H., Sparchez, Z.,

Serban, A., Suteu, T., and Badea, R. (2008). Toward

intelligent virtual biopsy: Using artiﬁcial intelligence

to predict ﬁbrosis stage in chronic hepatitis c patients

without biopsy. Journal of Hepatology, 48(2).

Freund, Y. and Schapire, R. E. (1997). A decision theoretic

generalization of on line learning and an application to

boosting. Journal of Computer and System Sciences,

55(1):119–139.

Guyon, I., Gunn, S., Nikravesh, M., and Zadeh, L. (2006).

Feature Extraction: Foundations and Applications.

Studies in Fuzziness and Soft Computing. Springer.

Lindor, A. (1996). The role of ultrasonography and

automatic-needle biopsy in outpatient percutaneous

liver biopsy. Hepatology, 23:1079–1083.

Poynard, T., Morra, R., Halfon, P., Castera, L., Ratziu, V.,

Imbert-Bismut, F., Naveau, S., Thabut, D., Lebrec, D.,

Zoulim, F., Bourliere, M., Cacoub, P., Messous, D.,

Muntenau, M., and de Ledinghen, V. (2007). Meta-

analyses of ﬁbrotest diagnostic value in chronic liver

disease. BMC Gastroenterology, 7(40).

Quinlan, J. (1993). C4.5 : Programs for Machine Learning.

Morgan Kaufmann.

Shaheen, A., Wan, A., and Myers, R. (2007). Fibrotest and

ﬁbroscan for the prediction of hepatitis c-related ﬁbro-

sis: a systematic review of diagnostic test accuracy.

Am J Gastroenterol, 102(11):2589–2600.

Tobkes, A. and Nord, H. J. (1995). Liver biopsy: Review of

methodology and complications. Digestive Disorders,

13:267–274.

INTELLIGENT CLINICAL DECISION SUPPORT SYSTEMS

287