Automatic extraction of numerical values from unstructured data in EHRs

Clinical data recorded in modern EHRs are very rich, although their secondary use research and medical decision may be complicated (eg, missing and incorrect data, data spread over several clinical databases, information available only within unstructured narrative documents). We propose to address the issue related to the processing of narrative documents in order to detect and extract numerical values and to associate them with the corresponding concepts (or themes) and units. We propose to use a CRF supervised categorisation for the detection of segments (themes, numerical sequences and units) and a rules-based system for the association of these segments among them in order to build semantically meaningful sequences. The average results obtained are competitive (0.96 precision, 0.78 recall, and 0.86 F-measure) and we plan to use the system with larger clinical data.