Information fusion for diagnosis coding support

All patient-related medical information during a hospital stay in France, has to be collected and coded in the compilation of medical units discharge documents, according to a standardized approach. The process of describing a patient disease in terms of appropriate diagnostic codes is nevertheless, a non-intuitive operation for the physician. As a consequence, coding errors, inaccuracies and missing data are frequent, leading to potentially severe economical upshots. A coding support system developed to improve medical coding results, integrates three information processing methodologies, using the outputs from various Hospital Information System applications. Each methodology generates partial heterogeneous information, with considerable semantic variety. In order to properly synthesize these outputs, information fusion is required to produce enriched contextualized information, presented to the physician as an ordered list of suggested codes. This paper explores two information fusion approaches: voting system and possibilistic. Both methods are tested on a database of 1,000 discharge summaries, to show the interest of information fusion in this context. Results show that fusion methods perform better in most of the cases than partial information extraction methods.