Introduction: Most of the European countries require indexing the hospital discharge abstract. Physicians and clinical coders, who are legally responsible for this task, need to get the simplest way to quickly retrieve the right code, especially in a decentralized coding environment. The context includes definitions, multiple labels, inclusion/exclusion notes and ICD10 tree structure position for the same disease. Although its richness, information contained in ICD10 hard-copy is accessed with difficulty in current practice. This project brings out associations between time and disease and between disease and aetiology, in order to enhance the disease context. The aim is to improve code retrieval and navigation be- tween all these notions using web technologies. Methods: iRMA data structure is based on a tree representation. Every node contains a code field associated with a text field. Moreover, for each node, an ordered set of linked information forms the siblings. Applied to ICD10, a node contains code, label and a set of linked information involving all analytical volume notes, alphabetical volume labels and local coding booklets. The same information can be linked to several nodes. Each informa- tion unit has its proper class in order to categorize meaning and presentation. The meaning of a node is formed by accumulating information from the root to the current position, following the branch. For example, information about a node with liver notifica- tion differs according to its location on the malignant neoplasm, benign neoplasm or intestinal infectious disease branch. The model is implemented as a relational database with Microsoft Access. As the classification tends to change quite rapidly, an interface allows its maintenance. This process can be either ma- nual updating or automated data importing. Files from the French official organization in charge of distributing classifications, "Agence Technique de l‟Information Hospitalière", are used as the source. Context building consists in looking for links between ICD10 codes. Studied relationships are on the one hand, disease and aetiology, and on the other hand, disease, sequellae and person- al/familial previous story. Most of the links for dagger and star codes are supplied by WHO-FIC. The other part is achieved by parsing text data with the help of PHP Programming language. We then produce a list of associations represented as couples, each of them being typed according to the studied relationship. Searching tool is implemented with HTML/PHP language. Two searching levels have been defined: the standard mode only looks for ICD10 analytical volume data; extended search looks for ICD alphabetical index and local data. Assessment of code retrieval was realized on the last real requests from physicians and clinical coders. An ICD code had been pre- viously attributed to each request by an expert coder. Each re- quest has been successively used as query terms for standard and extended search mode. The response was considered as a suc- cess if the expert‟s code was retrieved. The total number of re- turned codes was recorded. The mean recall and precision rates were then computed. Results: One hundred requests have been analysed. Mean num- ber of propositions per query was 1.1 ± 2.7 in standard mode and 2.9 ± 4.7 in extended mode (mean ± standard error). The query returned no code proposition for 52 and 12 requests respectively. The recall rate was 38% ± 49% in standard mode and 76% ± 43% in extended mode respectively. The precision rate was 60% ± 43% in standard mode and 51% ± 38% in extended mode. The interface is currently in use at Brest University Hospital, and also available on internet via: http://i3se009d.univ- brest.fr/IRMA/CIM10/CIM10.php. The user can browse ICD10 tree structure. He can also perform a search by single or multiple terms, and/or by code. The two search modes are proposed sepa- rately according to the user needs. Hits are displayed according to ICD10 hierarchy. Icons and colour list underline the code weight according to the French hospital payment system. By clicking on a code, its whole context is displayed. Hyperlinks allow switching from one context to another. Conclusions: This data model can be extended to other healthcare classifications. The web application shows encouraging results. First evaluations demonstrated the effectiveness of this method with a two times higher recall rate in extended method, despite a lower precision. Additional terms from local booklets improve appropriation by the local staff. Further developments are already planned, such as including other information sources, implement- ing a query follow-up tool and improving the code context with multilingual labels.