Modelling Cdss for Oncology

7/29/2019 Modelling Cdss for Oncology

1/8

International Journal of Medical Informatics (2005) 74, 299306

Modelling a decision-support system for oncology

using rule-based and case-based reasoningmethodologies

Delphine Rossillea,b,, Jean-Francois Laurentc, Anita Burguna

aLaboratoire dInformatique Medicale, Universite de Rennes 1, 2 avenue du Professeur Leon Bernard,

35043 Rennes Cedex, Franceb Department of Systems and Computer Engineering, Carleton University, Ottawa, Canadac Centre Eugene Marquis, Rennes, France

Received 31 October 2003; received in revised form 15 June 2004; accepted 23 June 2004

KEYWORDS

Information storage and

retrieval;

Models,

decision-support;

Decision-support

systems, clinical;

Data warehouse;

Medical informatics

applications;

Medical oncology

Summary In most hospital medical units, multidisciplinary committees meetweekly to discuss their patients cases. The medical experts base their decisionson three sources of information. First, they check if their patient complies withexisting guidelines. Failing these, the medical experts will base their therapeuticdecisions on the cases of similar patients that they have treated in the past. We

propose a multi-modal reasoning decision-support system based on both guidelineand case series, which will automatically compare the patients case to the cor-responding guideline, then to other cases, and retrieve similar cases. The generalstructure of the system is presented here, the domain of application being oncology.As the patients records are not currently stored in a database in a format which isdirectly accessible, an object-oriented model is proposed, which includes prognosisfactors currently tested in clinical trials, well-established ones, and a description ofthe illness episodes. The system is designed to be a data warehouse. Such a systemdoes not exist in the literature. Future work will be needed to define the similaritymeasures, and to connect the system to the current database. 2004 Elsevier Ireland Ltd. All rights reserved.

1. Introduction

In most hospital medical units, multidisciplinarycommittees (including surgeons, radiologists) meet

* Corresponding author. Tel.: +33 2 99 28 42 15;fax: +33 2 99 28 41 60.

E-mail address: [email protected] (D. Rossille).

weekly to discuss patients cases. The medical ex-perts base their decisions on three sources of in-formation: guidelines, clinical trials (either recentpapers discussing results, or ongoing clinical trials),and case series. Evidence-based medicine relies onthe guidelines issued and updated from the resultsof clinical trials. While most patients cases can beanalysed following these guidelines, some patients

1386-5056/$ see front matter 2004 Elsevier Ireland Ltd. All rights reserved.doi:10.1016/j.ijmedinf.2004.06.005


2/8

300 D. Rossille et al.

are non-standard (i.e. "atypical") because they donot conform to any guideline (e.g. rare illnesses), orthey do not comply with the entire guideline (e.g.the proposed treatment cannot be prescribed or itmust be aborted). For some of these cases, "close"enough to a guideline, the medical experts canadapt the guideline production rules to the cases.For some others, they will rely on clinical trials. Forthe remainder, any decision can only be based onsimilar cases the medical experts have encounteredin their career. Nowadays, most of the research is oncomputable rule-based reasoning decision-supportsystems (RBR DSS) (e.g. [13]). Systems consider-ing individual cases as a unique source of knowl-edge are becoming frequent in the medical domain(e.g. [4,5]). These case-based reasoning systems(CBR DSS) [6,7] follow four processes: retrieval ofsimilar cases, reuse of previous solutions applied tonew cases, evaluation of the proposed solution, and

case retaining.The system presented in this paper is meant to

be a data warehouse in oncology, storing valuableinformation for treating, for instance, patientswith rare tumours, or not reacting normally toa treatment. It will automatically retrieve sim-ilar cases, with a view to supporting medicalexperts when making decisions for non-standardcases, and to evaluate interpractice variationsby checking the consistency of decisions madefor similar cases. Nowadays, patients files arestored in large databases in formats not suitable

for automatic analysis. The first developmentphase of this research is to design the systemarchitecture and modelling. The second phasewill be concerned with the integration of thefunctionalities for searching and comparing similarcases.

The first phase is presented in this paper. UML(Unified Modelling Language) [8] is used as themodelling language. Because the system is basedon both rule-based reasoning (with guidelines) andcase-based reasoning (with individual cases), it is amulti-modal decision-support system. The system ispresented for breast cancer, and aims to be adapt-

able to any cancer. Its architecture was chosen sothat the system can evolve at the same pace asmedical knowledge.

2. Materials and methods

2.1. Materials

The system prototype was analysed on breastcancer, more specifically stage I invasive non-

metastatic breast cancer, and then the system wasgeneralized to any tumour.

The patients data come from the Centre EugeneMarquis (CEM), Rennes, France. The CEM storesaround 2000 new cases per year, one-fourth beingbreast cancers. The Centre is one of the 20 centresof the Federation Nationale des Centres de LutteContre le Cancer (FNCLCC) [9] and deals with one-fourth of all national cancers (50,000 new casesper year).

At the CEM, each patients medical record isstored both as a paper file and an electronic file.The electronic file is stored in the CEM proprietarydatabase partly as database fields and partly as doc-uments with no standardised format. Most patientsimages will shortly be stored in a separate PictureArchiving and Communication System (PACS).

The guidelines used by multidisciplinary commit-tees are described in the Standards, Options and

Recommendations (SOR) [9]. Domain-specific vo-cabularies and thesauri are used such as the ICD-O1

or the TNM2 classifications.The multidisciplinary committee for breast can-

cer currently discusses patients cases once a week.To identify the patient, the medical experts relyon a specific document called UCPS document gen-erated especially for the meeting. This documentsummarizes some of the most relevant medical dataas well as the history of the pathology for the pa-tient, and is used to record the unanimous decisiontaken by the committee. Whenever more medical

details are needed, the medical experts will con-sult the paper file of that patient. Indeed, the pa-per file is comprehensive as it includes documentsand images not yet digitalized (such as items of cor-respondence or mammograms), and as such it is themost suitable file on which the medical experts canbase their decision. On the other hand, the elec-tronic file already contains part of the relevant datain structured format suitable for automatic reason-ing. Todays decision making process for breast can-cer at the CEM is illustrated in Fig. 1. It shows whoplays a part in the committee meeting: the medicalexperts, their secretaries as well as the secretary

in charge of recording UCPS documents in the CEMdatabase.

2.2. Methods

The first step was to model todays decision makingprocess, which, even though based on breast cancer

1 ICD-O classification stands for International Classification ofDiseases for Oncology.

2 TNM classification stands for Tumour-Node-Metastases.


3/8

Modelling a decision-support system for oncology using RBR and CBR methodologies 301

cases, can be generalized to any type of tumours,and then to identify relevant data and knowledge.This was done in collaboration with the breast can-cer committee and based on the medical records ofthe CEM.

Because the decision-support system serves dif-ferent purposes from the CEM Hospital InformationSystem, we had to select the data relevant to ourobjectives.

For the purpose of this research, we selected(as indicated in Section 2.1) the stage I invasivenon-metastatic breast cancers, classified T1 N0 M0[10], diagnosed in 1997, having had or not a re-lapse since. This subset corresponds to pathologiesof normally easily treated cancers (tumour size lessthan 2 cm, no metastases and no nodes invasion).However, a significant number of these cancers mayevolve and some may have a fatal outcome. The di-agnosis date was chosen as far back as possible so

that any relapse could be recorded, whilst still pro-viding data recorded in an electronic format.

As the proposed system relies both on evidence-based medicine and experience-based medicine, asubset of all the patients data was chosen thatwould allow the comparison with the guidelines andsimilar cases, while containing suitable granularityfor supporting therapeutic decisions.

The DSS was then designed, and a class diagramof the medical cases was drawn using UML (Fig. 4).The choice of guidelines representation was deter-

Fig. 1 Todays decision making process for breast cancer at CEM.

mined by the fact that guidelines are structured andmust be computer-interpretable, as well as human-readable, in order to allow for automatic compari-son and viewing by experts.

3. Results

3.1. DSS architecture and reasoning process

The system consists of:

a base of individual cases, containing patientsrelevant medical data,

a knowledge base containing:o domain specific thesauri, including interna-

tional coding standards (ICD-O, TNM), nationalcoding standards, protocols for treatments,

medications;o a base of guidelines;o a classification table used by the reasoning

method in order to retrieve similar cases;o data used by the system administrator (re-

lated to users access rights for instance). a processing engine, a humancomputer interface.

The system is aimed to be fed by the HospitalInformation System (HIS) after appropriate format-ting of relevant information. Note however, that


4/8


Fig. 2 Sequence diagram of the use case "Searching for similar cases".

in most French HIS today this step would not bestraightforward as medical data are not collected

for the same purpose in the HIS as in our proposedsystem (for instance consultation reports written infree text contain relevant data that will have to besingled out to be fed to the proposed system).

The architecture separating the knowledge basefrom the programming code provides suitable flexi-bility in the system for any information update thatmight be made necessary by the rapid evolution ofmedical knowledge.

The humancomputer interface will be easy andstraightforward for non-computer experts such asdoctors or secretaries to use.

The computerized reasoning process goes as fol-

lows: It is hybrid being based on two reasoningmethodologiesevidence-based and experience-based medicine, and it is sequential.

Once a new case is stored in the cases base,the system will automatically select the appro-priate guideline and compare the new case withthis guideline. The selection of the guideline willbe made by comparing key medical terms (andtheir values) characterizing both medical cases andguidelines, these terms corresponding to commonrestricted vocabulary. For instance, for breast can-cer, based on the SOR for infiltrating non-metastatic

breast cancer, such data as the TNM classificationand the tumour size will be used. At the end of

the comparison of the new case with the guideline,the last guideline step with which the case com-plies will be saved in a classification table. Thanksto the classification table it will be possible to iden-tify directly which stored cases are classified underthe same guideline step. Note that the comparisonis only possible if the same vocabulary is used forboth cases and guidelines and if the values of themedical data are normalized, that is, if cases andguidelines "speak the same language with the samesemantic rules". This information would normallybe kept in the knowledge base.

When cases similar to a stored given case are

searched, the system will use the classification ta-ble to retrieve the identification numbers of all sim-ilar cases according to the guideline (Fig. 2). Foreach retrieved case, the system will then operatea more precise comparison on other predefined cri-teria in order to better select the closest similarcases to the given case.

3.2. Guidelines representation

The GLIF3 object model [11] was selected to rep-resent the structured guidelines. Unlike the Arden


5/8


Fig. 3 Use cases diagram of the proposed system.

syntax, with GLIF a guideline is described as aflowchart of temporally ordered steps, allowing thesystem to keep track of the step at which a case isno more compliant with the guideline. It also stores

information on the authors of the guidelines, theedition date, version, and eligibility criteria, whichwill make the selection of the appropriate guidelinein the guidelines base easier.

3.3. Patient data representation

3.3.1. Patients relevant data setA patients case becomes eligible to be stored inthe system only once the primary tumour diagno-sis has been made. This means that the databasereflects only the completed diagnosis, treatments

and examinations. Furthermore, a stored patientscase is active (that is, retrievable as a similar case)only if the patients state is known, either becausehe/she has given recent news3 (attribute recent-NewsDate in Fig. 4) or he/she has died (attributedate death).

Relevant patients data are of two types: prog-nosis factors and episodes of the disease history,

3 "Recent news" means "within the last year" as, once a year,the list of national yearly deaths is used to update the medicalrecords at the CEM.

an episode being a primary tumour diagnosis, or arecurrence, metastatic or none.

Prognosis factors are:

the well-established factors referenced in guide-

lines such as hormonal receptor levels (ER andPR) for breast cancer [9];

factors being currently validated in clinical trials(and not included yet in guidelines), for instancethe Her2/Neu marker [10];

factors that have modified the plan of care, suchas a treatment abortion and its reasons (the pa-tients refusal to continue the treatment or abor-tion due to harmful secondary effects).

These factors are either time-invariant andepisode-independent (as the mutation of genesBRCA1 and BRCA2), time-invariant but episode-

dependent (as ICD-OTTopography and ICD-OM Mor-phology codes for the primary tumour diagnosis),or time-variant and episode-dependent (as qualita-tive (qualitvalue) or quantitative (quantvalue) re-sults of exams). They are attached to the patientor to an episode according to their dependencies.

An episode is described by tumour-specific char-acteristics, associated treatments, and examina-tions. It includes temporal and factual data.Tumour-specific characteristics are of two kinds: ei-ther for primary tumour diagnosis, local and oppo-site mammary recurrences (as the ICD-O and TNM


6/8


codes), or for metastatic recurrence (as location).Treatments include Surgery, RT(radiotherapy), CT(chemotherapy), HT (hormonotherapy). The char-acteristics of each treatment are stored, includingsurgical procedure (e.g. mastectomy for Surgery),

protocols, duration or date of the treatment, re-sponse to treatment and all alarms occurring duringthe treatment. The order (rank) of the treatmentsis also stored. Medical images are included throughtheir qualitative results in the Exams class. The du-ration of an Episode (startDate being the date ofthe diagnosis and finalDate either the date of therecurrence or the final date of the last treatment)as well as a summary of the responses to treatmentsare provided.

The required granularity level of relevant datais, as a minimum, provided by the level used in theguidelines. In order to classify cases in relation tothemselves, additional details on relevant data will

be stored, such as protocols for chemotherapy andradiotherapy.

Patients data are stored essentially in struc-tured fields, in order to speed up cases retrieval.The fields format may be numerical values,dates, or restricted vocabulary. Restricted vo-cabulary includes cancer-specific vocabulary(e.g. episode type in Fig. 4 can be "primary tu-mour", "local recurrence", "opposite mammary

Fig. 4 Class diagram of the patient case.

recurrence", or "metastasis"), and coding stan-dards such as ICD-O. Commentary fields (such ascom episode) have been included to allow forexperts comments not included in other fieldsto be saved. Free text may be used in thesefields.

3.3.2. UML modellingThe actors of the system (Fig. 3) are the medicalexperts (i.e. surgeons and radiotherapists), the ad-ministrator and the CEM database. The administra-tor manages the system, including the guidelines(creation, suppression, update, activation) and theclassification process of the patients cases. Themedical experts consult a guideline or a patientscase, search for similar patients cases or check theconsistency of decisions for similar cases. The sys-tem cases base is fed by the CEM database.

The Patient Case class is represented in Fig. 4,

where the hierarchical structure can be seen.The latter comprises two branches: the patient-specific characteristics (the Patient class being itsroot), and the disease-specific characteristics (theEpisode class being its root). Note that only part ofthe attributes is shown in Fig. 4. In addition to thepatients relevant data listed in Section 3.3.1, eachcase is related to the Classification class that storesthe step (leaf or node of the associated guideline


7/8


or decision tree) at which the case does not complywith the guideline anymore, and henceforth, linkstogether similar cases found after the first classifi-cation according to the guideline.

In order for the system to be suitable to anal-yse any tumour, the case has an object-oriented ar-chitecture composed of generic classes (including

patient, Alarms, FamilyHistory, Episode, Charac-teristicsCancer, CharacteristicsMetastasis, Treat-ments (CT, RT, HT, Surgery), Exams), to whichcancer-dependent classes are associated or in-herited (such as BreastCancer characterizing theprimary breast tumour, or BreastFactors charac-terizing the episode-independent breast cancer-dependent factors). In addition, for some genericclasses (such as RT, CT, HT, Surgery), the contentis cancer-dependentthat is, an attribute can onlytake values specific to a given cancer (the proto-col for chemotherapy for breast cancer is different

from that for liver cancer).

4. Discussion

Most of the medical decision-support systems (DSS)rely nowadays on a unique reasoning methodologysuch as case-based reasoning (CBR) or rule-basedreasoning (RBR). Yet, a few DSS base their reason-ing process on a multi-modal approach. The systemproposed in this paper is a multi-modal DSS basedon both CBR and RBR. Four different approaches

combining cases and rules may be found in the liter-ature. The hybrid RBR-first CBR-last approach usesthe RBR as its main reasoning process, CBR beingapplied when RBR is not found suitable that is forexceptions to the rules and/or non-standard situa-tions [12]. With the CBR-first RBR-last approach, theCBR is the main reasoning process and RBR is usedto improve part of the reasoning process [13]. Othersystems apply CBR and RBR in parallel. Either bothoutcomes are simply displayed [14], or the bestoutcome is proposed according to some given cri-teria. The CARE-PARTNER system on bone-marrowand stem-cell post-transplant long-term follow-up

[15] implements a closer cooperation between thetwo methodologies, through the iteration of partialreasoning steps, the outcome of each of these stepsbeing the best outcome of the two methodologies.Yet, all of the above-mentioned multi-modal sys-tems apply CBR and RBR in mutually exclusive ways.The T-IDDM project [16] on diabetes managementappears to be a multi-modal system realizing a realintegration of both CBR and RBR. Indeed, CBR re-sults are used to refine the rules, hence allowing thetailoring the final outcome on the specific patientsneeds. Kasimir [17], applied to cancer cases, is sim-

ilar as it uses case-based reasoning to adapt theproduction rules of the guideline to atypical cases.

In oncology, the trend is for DSS systems to relyon the CBR approach [4,5]. The proposed system is ahybrid RBR-first CBR-last system, aiming at support-ing the medical experts in their decisions for non-standard (atypical) cases. In the literature, onlyKasimir [17] is a similar multi-modal system ap-plied to cancer, either for standard cases or non-standard cases close to standard ones. However theproposed system has a different and complemen-tary approach as it can be useful even when noguideline exists for a specific rare disease.

The paper presents a patient data model forthe multi-modal reasoning system. Several or-ganisations (including HL7, CEN/TC251, openEHRFoundation) have been working on defining anobject-oriented structure of hospital ElectronicPatient Records (EPR) [2]. They provide rep-

resentation models of patient data (includingclinical data), and domain-specific concepts, andinclude data exchange capability (within thehospital as well as with outside partners). Theyare both intended for any health domain. The HL7RIM appears too general to be directly usable indecision-support systems [2]. The CEN/TC251 EHRCand openEHR GEHR, as well as, for guideline-basedsystems, the EON and PRODIGY architectures [1,2],contain data which are not relevant to the pro-posed system as they correspond to non-completedevents (e.g. goals, planned interventions). The

proposed architecture is close to the openEHRGEHR: the persistent transactions correspond-ing to the patient branch, and process-selectedevent-driven transactions to the Episode branch.However, the openEHR GEHR is designed to includeall the aspects of the patient medical record withinthe Hospital Information System (as HL7 RIM andCEN/TC251 EHRC), while the proposed systemrelies on an independent cases base. In CBR DSSapplied to oncology, Kasimir [17] bases its decisionsonly on the clinical data required by the guidelines,while the proposed system includes potential fac-tors as well. Other CBR systems for breast cancer

represent very specific information on the patient(e.g. histo-pathological data in [4]), whereas theproposed system is applicable to any tumour. UMLwas shown to be efficient to model the system andconfirms its applicability to CBR systems as wasdemonstrated in other studies, e.g. [18].

The proposed patient case may model anytype of cancers. Not all cancers, however, haveguidelines as structured as breast cancer andtheir automation can become more delicate. Thisis a very common difficulty encountered whendesigning computable guidelines. No real solution


8/8


can be found from reading the literature, apartfrom the suggestion of interactively asking theexpert to validate a choice [3]. This solution is notapplicable to our system. For the future system tobecome entirely automated, patients data shouldbe automatically acquired from the CEM database.However, most of the required data are still in elec-tronic documents at the CEM, even though part ofthem is stored in structured fields (such as the ICD-O). Relevant data will thus have to be first identifiedin electronic documents and formatted appropri-ately before being stored in the proposed system.

Finally, the proposed system is a data warehousewhich is flexible enough to be able to integrate im-ages in the future (including data from functionalimagery of tumours) as well as new data describedby emerging disciplines (factors such as genes ex-pressed in cancers, or raw data such as patientsmicro-array results).

5. Conclusion

This paper presents the modelling of a case-basedretrieval which is designed to support therapeuticdecisions in cases that do not or cannot comply withrecommended guidelines. No such system exists asyet in the literature. This paper detailed the gen-eral architecture of the system and the modellingof patients cases and guidelines. The next researchstage will be to define the similarity measures tobe applied to patients data in order to allow theretrieval of similar cases and also to design a user-friendly human/machine interface. The proposedsystem is flexible enough for new data to be in-corporated (for instance from emerging disciplinessuch as molecular biology) and should be able totake into account the rapid evolution of medicalknowledge (e.g. [19]).

References

[1] S.W. Tu, A.M. Musen, Modeling data and knowledge in theEON guideline architecture, MEDINFO (2001) 280284.

[2] P. Johnson, S.W. Tu, A.M. Musen, I. Purves, A virtual medi-cal record for guideline-based decision support, Proc. AMIASymp. (2001) 294300.

[3] B. Seroussi, J. Bouaud, E. Antoine, ONCODOC: a successfulexperiment of computer-supported guideline developmentand implementation in the treatment of breast cancer, Ar-tif. Intell. Med. 22 (1) (2001) 4364.

[4] M.-C. Jaulent, C. Le Bozec, E. Zapletal, P. Deguoulet, Casebased diagnosis in histopathology of breast cancer, MEDINFO(1998) 544548.

[5] R. Schmidt, L. Gierl, Case-based reasoning for antibioticstherapy advice: an investigation of retrieval algorithms andprototypes, Artif. Intell. Med. 23 (2) (2001) 171186.

[6] A. Aamodt, E. Plaza, Case-based reasoning: foundationalissues, methodological variations, and system approaches,

in: AI Communications, IOS Press, 1994.[7] J. Kolodner, Case-based Reasoning, Morgan Kaufmann, San

Francisco, CA, 1993.[8] P.-A. Muller, N. Gaertner, Modelisation objet avec UML, 2nd

ed., Eyrolles, 2000.[9] Federation Nationale desCentres de Lutte Contre le Cancer,

SOR: Cancers du sein infiltrants non metastatiques, 2nd ed.,John Libbey Eurotext, 2001. www.fnclcc.fr.

[10] S. Singletary, C. Allred, P. Ashley, L. Bassett, D. Berry, K.Bland, P. Borgen, G. Clark, P. Edge, D. Hayes, L. Hughes, R.Hutter, M. Morrow, D. Page, A. Recht, R. Theriault, A. Thor,D. Weaver, S. Wieand, F. Greene, Revision of the AmericanJoint Committee on cancer staging system for breast can-cer, J. Clin. Oncol. 20 (17) (2002) 36283636.

[11] M. Peleg, A. Boxwala, O. Ogunyemi, Q. Zeng, S. Tu, R. Lac-

son, E. Bernstam, N. Ash, P. Mork, L. Ohno-Machado, E.Shortliffe, R. Greenes, GLIF3: the evolution of guidelinerepresentation format, Proc. AMIA Symp. (2000) 645649.

[12] H.-J. Park, J.-S. Oh, D.-U. Jeong, K.-S. Park, Automatedsleep stage scoring using hybrid rule- and case-based rea-soning, Comput. Biomed. Res. (33) (2000) 330349.

[13] G.J. Petot, C. Marling, L. Sterling, An artificial intelligencesystem for computer-assisted menu planning, J. Am. Di-etetic Assoc. (1998) 10091014.

[14] K. Evans-Romaine, C. Marling, Prescribing exercice regi-mens for cardiac and pulmonary disease patients with CBR,ICCBR (2003).

[15] I. Bichindaritz, M. Siadak, J. Jocom, C. Moinpour, E. Kansu,G. Donaldson, N. Bush, M. Chapko, J. Bradshaw, K. Sullivan,CARE-PARTNER: a computerized knowledge-support systemfor stem-cell post-transplant long-term follow-up on theworld wide web, Proc. AMIA Symp. (1998) 386390.

[16] S. Montani, R. Bellazzi, Supporting decisions in medical ap-plications: the knowledge management perspective, Int. J.Med. Inform. 68 (2002) 7990.

[17] J. Lieber, M. DAquin, P. Bey, B. Bresson, O. Croissant, P.Falzon, A. Lesur, J. Leveque, V. Mollo, A. Napoli, M. Rios,C. Sauvagnac, The Kasimir project: knowledge manage-ment in cancerology, in: Proceedings of the Fourth Inter-national Workshop on Enterprise Networking and Comput-ing in Health Care Industry (HealthComm 2002), 2002, pp.125127.

[18] C. LeBozec, M.-C. Jaulent, E. Zapletal, P. Degoulet, Unifiedmodeling language and design of a case-based retrieval sys-tem in medical imaging, Proc. AMIA Symp. (1998) 887891.

[19] M.J. Van de Vijver, Y.D. He, L.J. vant Veer, H. Dai, A.A.Hart, D.W. Voskuil, G.J. Schreiber, J.L. Peterse, C. Roberts,M.J. Marton, M. Parrish, D. Atsma, A. Witteveen, A. Glas,L. Delahaye, T. van der Velde, H. Bartelink, S. Rodenhuis,E.T. Rutgers, S.H. Friend, R. Bernards, A gene-expressionsignature as a predictor of survival in breast cancer, N. Engl.J. Med. 347 (25) (2002) 19992009.
http://www.fnclcc.fr/http://www.fnclcc.fr/

Modelling Cdss for Oncology

Documents

Transcript of Modelling Cdss for Oncology