Using MetaMap to extract medical concepts from unstructured clinical notes written in Italian

We assessed the feasibility of using MetaMap to extract disorders, symptoms and findings from Italian clinical notes. We performed two experiments: in "Exp A" we used MetaMap and a knowledge source consisting of Italian UMLS sources to annotate Italian clinical notes, in "Exp B" we used MetaMap to analyze an English unsupervised translated version of the original Italian texts. Average recall, precision and Fmeasure were equal to 0.53, 0.98 and 0.69, in "Exp A", and to 0.75, 0.95 and 0.83, in "Exp B". For both experiments, MetaMap showed better performances for "Disorders" than for "Findings" and "Symptoms". In conclusion, MetaMap's performances with unstructured clinical notes written in Italian could be improved modifying MetaMap to allow generating Italian variants. MetaMap's performances with English translation the original Italian texts, performances were good enough to allow using MetaMap in clinical practice. Further improvements could be obtained using a supervised translation of medical terms.