BACKGROUND: Medical plaintext documents contain important facts about patients, but they are rarely available for structured queries. The provision of structured information from natural language texts in addition to the existing structured data can significantly speed up the search for fulfilled inclusion criteria and thus improve the recruitment rate. OBJECTIVES: This work is aimed at supporting clinical trial recruitment with text mining techniques to identify suitable subjects in hospitals. METHOD: Based on the inclusion/exclusion criteria of 5 sample studies and a text corpus consisting of 212 doctor's letters and medical follow-up documentation from a university cancer center, a prototype was developed and technically evaluated using NLP procedures (UIMA) for the extraction of facts from medical free texts. RESULTS: It was found that although the extracted entities are not always correct (precision between 23% and 96%), they provide a decisive indication as to which patient file should be read preferentially. CONCLUSION: The prototype presented here demonstrates the technical feasibility. In order to find available, lucrative phenotypes, an in-depth evaluation is required.
BACKGROUND: Medical plaintext documents contain important facts about patients, but they are rarely available for structured queries. The provision of structured information from natural language texts in addition to the existing structured data can significantly speed up the search for fulfilled inclusion criteria and thus improve the recruitment rate. OBJECTIVES: This work is aimed at supporting clinical trial recruitment with text mining techniques to identify suitable subjects in hospitals. METHOD: Based on the inclusion/exclusion criteria of 5 sample studies and a text corpus consisting of 212 doctor's letters and medical follow-up documentation from a university cancer center, a prototype was developed and technically evaluated using NLP procedures (UIMA) for the extraction of facts from medical free texts. RESULTS: It was found that although the extracted entities are not always correct (precision between 23% and 96%), they provide a decisive indication as to which patient file should be read preferentially. CONCLUSION: The prototype presented here demonstrates the technical feasibility. In order to find available, lucrative phenotypes, an in-depth evaluation is required.