A Web application to extract expert
on natural history of diseases.
Alain-jacques Valleron (1), Jean-François Vibert (1), Bernard Larouzé (1), Sylvie Deuffic-Burban (1), Mostafa K. Mohamed (2), Mokhtar Dalichaouche (3) and Jacques Vidal (4)
(1) Epidémiologie et Sciences de l’Information, INSERM U444, Hôpital Saint-Antoine. Paris, France.
(2) Department of Community, Faculty of Medicine, Aïn Shams University, Abbassia, Cairo, Egypt
(3) Service des Maladies Infectieuses, CHU et Faculté de Médecine de Constantine, Constantine, Algérie
(4) Computer Science Department, UCLA, Los Angeles, California, USA
We describe a method devised to draw, from a distributed expertise, the information necessary to model the epidemiology of a disease for which little data is available. The dynamic Web based application which has been developed (DSW, Delphi_Survey_Web) is a fully configurable application adaptable to many applications.
Delphi survey, web application, expert knowledge extraction,
Epidemiologic modeling allows to quantify the burdens of disease and to anticipate the cost/benefits of preventive program and/or of new treatments. It is a key resource in three domains: emergent diseases, bioterrorism and health effects of exposure to very low doses of chemical or physical agents. The purpose of this work is to describe a web tool designed to extract the knowledge necessary to the setting of the models, in absence of field data.
A first example of application was the epidemiology of HCV infection in Mediterranean countries. In developed countries, the frequency of HCV is less then 1%. In south Mediterranean countries the frequency is considerably higher (30% in some regions of Egypt) and the disease may be different as the genotype of the virus and the cofactors of the liver disease progression are different. While actual epidemiological data is scarce, there are good experts.
Each problem ("Survey") is under the responsibility of a principal investigator (PI). When the studies are multicentric, each center is under the responsibility of a “coordinator”. Each coordinator is in charge of a group of experts. Diseases are modeled as a multistage process. Patients may progress through n stages. Among all the n x n possible transitions between any two stages, those who are plausible are specified by the model builder (the PI). The transition probabilities per year may be governed by cofactors, such as age, sex, genetic characteristics, etc. Each cofactor can be text, value, range or binary choice. Quantitative cofactors (e.g. age) are categorized in classes. A series of case histories is then automatically generated for all possible combinations of transitions (<10), of cofactors (<5) per transition and of levels of the cofactors. In the HCV example, this leads to 5 x 3 x 3 = 45 cases. This sample may be repeated r times (in the HCV example, r =2). The PI can edit the case stories and add clinical details. The final design of experiment is therefore a factorial design with r repetitions. This allows a statistical analysis testing the perceived role of each cofactor. The experts receive cases histories and questions such as “this is a patient presenting a liver cirrhosis. Male, Age 50, likely infected by HCV 20 years before. Give your estimate of the probability that this subject will die during the next year”. As each round, the answers are analyzed and the probabilities are computed on line.
Acknowledgments: This work is supported by a grant of the European Union (VHSMC).
Address for correspondence:
Hôpital Saint-Antoine, Assistance Publique - Hôpitaux de Paris
182, rue du Faubourg Saint Antoine - 75012 PARIS FRANCE