Task 1a of the CLEF eHealth Evaluation Lab 2015

Best practice for clinical handover and its documentation recommends standardized, structured, and synchronous processes with patient involvement. Cascaded speech recognition (SR) and information extraction could support their compliance and release clinicians' time from writing documents to patient interaction and education. However, high requirements for processing correctness evoke methodological challenges. First, multiple people speak clinical jargon in the presence of background noise with limited possibilities for SR personalization. Second, errors multiply in cascading and hence, SR correctness needs to be carefully evaluated as meeting the requirements. This overview paper reports on how these issues were addressed in a shared task of the eHealth evaluation lab of the Conference and Labs of the Evaluation Forum in 2015. The task released 100 synthetic handover documents for training and another 100 documents for testing in both verbal and written formats. It attracted 48 team registrations, 21 email confirmations, and four method submissions by two teams. The submissions were compared against a leading commercial SR engine and simple majority baseline. Although this engine performed significantly better than any submission [i.e., 38.5 vs. 52.8 test error percentage of the best submission with the Wilcoxon signed-rank test value of 302.5 (p < 10-12)], the releases of data, tools, and evaluations contribute to the body of knowledge on the task difficulty and method suitability.