DIAGNOSIS OF HUMAN EPENDYMOMAS: COMPUTER-ASSISTED DECISION-MAKING WITH THE BAYESIAN TOOL "MEDES"  

Article type: Research Article

Authors: Iglesias, José R. | Esparza, Javier | Sanchez, Beatriz | Strik, Herwig

Affiliations: Institut für Pathologie, (Neuropathologie), Katharinenhospital, Postfach 10 26 44, D-70022 Stuttgart. Germany. Tlf: +49 711 278 4918. Fax: +49 711 2784909 e-mail: jr.iglesias@katharinenhospital.de | Institut für Informatik, Technische Universität München, Arcisstr. 21, D-80290 München, Germany. Tlf: +49 89 2892 2405. Fax: +49 89 2892 8207 E-mail: {esparza,sanchez}@informatik.tu-muenchen.de | Institut für Hirnforschung, Eberhard-Karls-Universität Tübingen, Calwer Str. 3, D-72076 Tübingen, Germany E-Mail: hgstrik@med.uni-tuebingen.de

Note: [] Address for correspondence: Priv.-Doz. Dr. med. José R. Iglesias Institut für Pathologie, (Neuropathologie) Katharinenhospital Postfach 10 26 44 D-70022 Stuttgart.Germany. E. mail: jr.iglesias@katharinenhospital.de and iglesias@z.zgs.de  

Abstract: Diagnostic decision-making in pathology involves gathering observational evidence and using it to reason and to decide which out of a number of possible diagnoses is most likely. In many areas of Pathology, a rich body of objective morphological numerical information exists. However, pathologists cannot make optimal use of this information, because, like all humans, they are not good at handling large amounts of numerical data. MEDES (Medical Decision Shell), a tool developed by the authors, has been used to develop a decision-making model for histological and clinical diagnosis of human ependymomas. 268 human ependymomas collected in our laboratory of neuropathology were studied. 176 ependymomas were used as training test: 79 typical ependymomas, WHO-Grade II), 34 subependymomas, 22 myxopapillary ependymomas, and 41 anaplastic ependymomas. The frequencies of location, age and sex and of forty one histological features were considered. All histological features were evaluated on a four-degree scale. On the basis of this information, MEDES produces a decision making model using Bayesian statistics. 92 new cases of ependymomas were then tested. MEDES selects the question that discriminates best between the diagnoses, updates the probabilities of the diagnosis according to the answer with Bayesian statistics, and iterates the procedure. MEDES stops asking questions when further evidence cannot change the most probable diagnosis. With this method, 91.3% of the tested tumors were classified correctly. MEDES allows to make objective use of statistical data in the histological diagnosis of ependymomas. It discriminates between important and unimportant questions, and asks only for the relevant information. MEDES provides pathologists and students with a vehicle to improve logical consistency and reproducibility in diagnostic practice and objectivity of diagnostic decisions.

Keywords: decision making, Bayesian Network, Glioma, Ependymoma, diagnosis

Journal: Electronic Journal of Pathology and Histology, vol. 7, no. 3, pp. 08-08, 2001