Finding Patient Visits in EMR Using LUXID

INTRODUCTION : Free text sections of the Electronic Medical Records (EMR) contain information that cannot be appropriately constrained in the structured forms. Several studies have shown the potential utility in mining EMR free texts for identifying adverse events (e.g. EU-PSIP, EU-ALERT), and large public-private research projects (e.g. IMI-EHR4CR, CLOUD4HEALTH) aim at mining them further, e.g. for clinical trial optimisation and pharmacovigilance purposes. AIM : The purpose of this work has been to assess the performance of LUXID R ©, an o -the-shelve commercial natural language processing system, using the dictionaryand rule-based Medical Entity Relationships Skill Cartridge R ©and KNIME as automation work ow engine for result combination and formatting, on the University of Pittsburgh BLULab NLP Repository benchmark, in the context of the TREC 2011 Medical Records Retrieval Track (TREC-MED2011). RESULTS : The system here described achieved the best score for one of the 34 queries (de ned as query 111) and overall classi ed as top 7th-8th (according to the scoring used) in the manual track of TREC-MED2011. More than 80% of the queries of TREC-MED2011 could be appropriately processed automatically. Performance of manually interpreted queries did not di er substantially from those automatically processed. More than 60% of the queries submitted by our system delivered a performance above or on the median of all participants. Very high precision of the system, delivering in certain cases a very low number of hits, correlated statistically with the overall performance. CONCLUSIONS : Initial results, error analysis are reported and strategies for improvements of the system are outlined; fully supporting the appropriateness in using this technology for identifying patients matching inclusion/exclusion criteria using plain text from unstructured EMR.