BT-45: Automatic Textual Summarisation of Neonatal Intensive Care Data

As ICUs generate increasing amounts of information, writing medical reports involves complex time-consuming reasoning to build a coherent text which will be meaningful to those who will use it for decisions making (e.g.: for nurse handover). To facilitate this summarisation, we developed a system called BT-45 that automatically generates textual summaries from periods of continuous and discrete data in a neonatal ICU. The system is related to the BabyTalk project [1] whose aim is to support clinical staff in a neonatal ICU by providing a textual summary of the state of a baby over a period of time; and to facilitate the writing of medical reports. The demonstration will show the system running on real data and will detail the steps of the processing. The BT-45 prototype is an extension of the NLG pipeline architecture originally proposed in [2]. It automatically summarizes 45 minutes of continuous and discrete data in four stages. Physiological signals (heart rate, etc.) are processed by a Signal Analysis module to abstract the main features of the signals. A Data Abstraction module infers relations (causality, association, etc.) between signals features and other discrete events (lab results, medication, etc.) and abstracts patterns of events into higher description levels. A Content Determination module then selects the most important events and features, and aggregates them into a tree of linked events. Finally, the Micro Planning module translates this tree into coherent text. The modules are linked using a Protégé-Frame ontology for the NICU of concepts acquired from clinicians. Few NLG systems consist of such a complex processing chain that goes from the signals to meaningful text. BT-45 has been tested in a clinic trial. Over thirty nurses and doctors were asked (individually) to say what clinical actions(s) they would take for a baby whose recent history over a period of about 45 minutes was presented either graphically (G), as text generated by a human expert (H) or as text generated by BT-45 (C). The appropriateness of the action chosen served as score for evaluating each mode of presentation. Result showed that H leads to significantly better decisions than with G and that our computer system C performed at least as well as G (no significant difference). After only one year of development, this result encourages us to develop extended systems, specifically tailored to doctors, nurses and family of the newborn.