24 June 2010
Automatically generating shift summaries in a Neonatal Intensive Care Unit
Meeting Room 10, 2nd Floor, JLB
12:30pm - 13:45pm
Dr Albert Gatt - University of Malta
This talk describes the design and evaluation of BT-Nurse, a Natural Language Generation (NLG) system built as part of the BabyTalk Project and recently evaluated on ward in a Neonatal Intensive Care Unit. The system automatically generates a summary from the raw data collected about a patient in the course of a shift, giving an overview of a baby’s current status, as well as “narrating” the clinically important events that have occurred. It was built in an effort to provide better support for decision-making and care planning by medical staff on the ward, while alleviating some of the burden of writing shift summary reports.
An earlier evaluation of a prototype system showed that NLG technology can support decision-making at least as well as existing technologies for information presentation on the ward. However, this early prototype was limited in that it summarized a small amount of data (45 minutes), some of which had been manually pre-processed. Moreover, the evaluation was conducted off-ward in controlled conditions.
A clearer demonstration of the effectiveness of NLG technology would require considerable scaling up to deal with very large volumes of noisy, heterogeneous data collected over a 12-hour shift. Apart from dealing with the sheer volume of data, the generation of effective summaries for decision support also needs to take into account a number of linguistic issues in order to produce a fluent, coherent, narrative that allows an informed reader to identify the clinically important events related to a patient, as well as their temporal and causal relationships.
BT-Nurse was developed to meet these requirements, but it also involved a step beyond existing evaluation practices in NLG, in that it was deployed on ward for a period of several months and used under supervision by incoming and outgoing nurses, whose feedback on various aspects of the system was collected and is currently being analysed. One of the aims of this talk is to identify some of the key lessons learned from this evaluation exercise.
Keywords: natural language generation (NLG), medical decision support, narrative, evaluation
Save to your Calendar