Les effets indésirables liés aux médicaments (EIM) sont des dommages liés au traitement médicamenteux plutôt qu’aux conditions sous-jacentes du patient. Ils mettent les patients en danger, et la plupart d’entre eux sont évitables. La détection des EIM repose habituellement sur les reports spontanés d’EIM et sur la revue de dossiers. L’objectif du présent travail est d’identifier automatiquement les cas d’EIM en utilisant des méthodes de Data Mining (fouille statistique de données). Le DataMining est un ensemble de méthodes statistiques particulièrement adaptées à la découverte de règles dans de grandes bases de données.Matériel Un modèle de données commun est tout d’abord défini, dans le but de décrire les données qui peuvent être extraites des dossiers patient électroniques. Plus de 90 000séjours hospitaliers complets sont extraits de 5 hôpitaux français et danois. Ces enregistrements incluent les diagnostics, les résultats de biologie, les médicaments administrés, des informations démographiques et administratives, et enfin du texte libre (courriers, comptes-rendus). Lorsque les médicaments ne peuvent être extraits d’un CPOE (système de prescription connectée), ils sont extraits des courriers pa rSemantic Mining (fouille de texte). De plus, la société Vidal fournit un ensemble exhaustif de RCP (Résumés des Caractéristiques du Produit).Méthode On tente de tracer dans les données tous les événements indésirables décrits dans les RCP. Puis en utilisant les méthodes de Data Mining, en particulier les arbres de décision et les règles d’association, on identifie les circonstances qui favorisent l’apparition d’EIM. Plusieurs règles de détection des EIM sont ainsi obtenues, elles sont ensuite filtrées et validées par un comité d’experts. Enfin, les règles sont décrites sous forme de fichiers XML et stockées dans une base. Elles sont exécutées afin de calculer certaines statistiques et de détecter les cas d’EIM.Résultats236 règles de détection des EIM sont ainsi découvertes. Elles permettent de détecter27 types d’événements indésirables différents. Plusieurs statistiques sont calculées automatiquement pour chaque règle dans chaque service, comme la confiance ou le risque relatif. Ces règles impliquent des conditions innovantes : par exemple certaines règles décrivent les conséquences de l’arrêt d’un médicament. De plus, deux outils Web sont développés et mis à la disposition des praticiens via Internet : les Scorecards permettent de présenter des informations statistiques e tépidémiologiques sur les EIM propres à chaque service, tandis que l’Expert Explorer permet aux médecins d’examiner en détail les cas probables d’EIM de leur service.Enfin, une évaluation préliminaire de l’impact clinique des EIM est menée, ainsi que l’évaluation de la précision de détection des EIM. / Adverse Drug Events (ADE) are injuries due to medication management rather than the underlying condition of the patient. They endanger the patients and most of them could be avoided. The detection of ADEs usually relies on spontaneous reporting ormedical chart reviews. The objective of the present work is to automatically detectcases of ADEs by means of Data Mining, which are a set of statistical methodsparticularly suitable for the discovery of rules in large datasets.MaterialA common data model is first defined to describe the available data extracted from the EHRs (electronic health records). More than 90,000 complete hospital stays areextracted from 5 French and Danish hospitals. Those complete records includediagnoses, lab results, drug administrations, administrative and demographic data aswell as free-text reports. When the drugs are not available from any CPOE(Computerized Prescription Order Entry), they are extracted from the free-text reports by means of semantic mining. In addition, an exhaustive set of SPCs (Summaries of Product Characteristics) is provided by the Vidal Company.MethodsWe attempt to trace all the outcomes that are described in the SPCs in the dataset. By means of data mining, especially Decision Trees and Association Rules, the patternsof conditions that participate in the occurrence of ADEs are identified. Many ADEdetection rules are generated; they are filtered and validated by an expert committee. Finally, the rules are described by means of XML files in a central rules repository, and are executed again for statistics computation and ADE detection.Results236 ADE-detection rules have been discovered. Those rules enable to detect 27different kinds of outcomes. Several statistics are automatically computed for eachrule in every medical department, such as the confidence or the relative risk. Thoserules involve innovative conditions: for instance some of them describe theconsequences of drug discontinuations.In addition, two web tools are designed and are available through the web for thephysicians of the departments: the Scorecards enable to display statistical andepidemiological information about ADEs in a given department and the ExpertExplorer enables the physicians to review the potential ADE cases of theirdepartment.Finally, a preliminary evaluation of the clinical impact of the potential ADEs isperformed as well as a preliminary evaluation of the accuracy of the ADE detection.

Adverse Drug Events (ADE) are injuries due to medication management rather than the underlying condition of the patient. They endanger the patients and most of them could be avoided. The detection of ADEs usually relies on spontaneous reporting ormedical chart reviews. The objective of the present work is to automatically detectcases of ADEs by means of Data Mining, which are a set of statistical methodsparticularly suitable for the discovery of rules in large datasets.MaterialA common data model is first defined to describe the available data extracted from the EHRs (electronic health records). More than 90,000 complete hospital stays areextracted from 5 French and Danish hospitals. Those complete records includediagnoses, lab results, drug administrations, administrative and demographic data aswell as free-text reports. When the drugs are not available from any CPOE(Computerized Prescription Order Entry), they are extracted from the free-text reports by means of semantic mining. In addition, an exhaustive set of SPCs (Summaries of Product Characteristics) is provided by the Vidal Company.MethodsWe attempt to trace all the outcomes that are described in the SPCs in the dataset. By means of data mining, especially Decision Trees and Association Rules, the patternsof conditions that participate in the occurrence of ADEs are identified. Many ADEdetection rules are generated; they are filtered and validated by an expert committee. Finally, the rules are described by means of XML files in a central rules repository, and are executed again for statistics computation and ADE detection.Results236 ADE-detection rules have been discovered. Those rules enable to detect 27different kinds of outcomes. Several statistics are automatically computed for eachrule in every medical department, such as the confidence or the relative risk. Thoserules involve innovative conditions: for instance some of them describe theconsequences of drug discontinuations.In addition, two web tools are designed and are available through the web for thephysicians of the departments: the Scorecards enable to display statistical andepidemiological information about ADEs in a given department and the ExpertExplorer enables the physicians to review the potential ADE cases of theirdepartment.Finally, a preliminary evaluation of the clinical impact of the potential ADEs isperformed as well as a preliminary evaluation of the accuracy of the ADE detection.