Evaluation of natural language processing from emergency department computerized medical records for intra-hospital syndromic surveillance

BACKGROUND The identification of patients who pose an epidemic hazard when they are admitted to a health facility plays a role in preventing the risk of hospital acquired infection. An automated clinical decision support system to detect suspected cases, based on the principle of syndromic surveillance, is being developed at the University of Lyon's Hôpital de la Croix-Rousse. This tool will analyse structured data and narrative reports from computerized emergency department (ED) medical records. The first step consists of developing an application (UrgIndex) which automatically extracts and encodes information found in narrative reports. The purpose of the present article is to describe and evaluate this natural language processing system. METHODS Narrative reports have to be pre-processed before utilizing the French-language medical multi-terminology indexer (ECMT) for standardized encoding. UrgIndex identifies and excludes syntagmas containing a negation and replaces non-standard terms (abbreviations, acronyms, spelling errors...). Then, the phrases are sent to the ECMT through an Internet connection. The indexer's reply, based on Extensible Markup Language, returns codes and literals corresponding to the concepts found in phrases. UrgIndex filters codes corresponding to suspected infections. Recall is defined as the number of relevant processed medical concepts divided by the number of concepts evaluated (coded manually by the medical epidemiologist). Precision is defined as the number of relevant processed concepts divided by the number of concepts proposed by UrgIndex. Recall and precision were assessed for respiratory and cutaneous syndromes. RESULTS Evaluation of 1,674 processed medical concepts contained in 100 ED medical records (50 for respiratory syndromes and 50 for cutaneous syndromes) showed an overall recall of 85.8% (95% CI: 84.1-87.3). Recall varied from 84.5% for respiratory syndromes to 87.0% for cutaneous syndromes. The most frequent cause of lack of processing was non-recognition of the term by UrgIndex (9.7%). Overall precision was 79.1% (95% CI: 77.3-80.8). It varied from 81.4% for respiratory syndromes to 77.0% for cutaneous syndromes. CONCLUSIONS This study demonstrates the feasibility of and interest in developing an automated method for extracting and encoding medical concepts from ED narrative reports, the first step required for the detection of potentially infectious patients at epidemic risk.