Summarizing Neonatal Time Series Data

Summarizing Neonatal Time Series Data Somayajulu G. Sripada, Ehud Reiter, Jim Hunter and Jin Yu Department of Computing Science University of Aberdeen, Aberdeen, U.K. fssripada,ereiter,jhunter,jyul@csd.abdn.ac.uk Abstract We describe our investigations in gener- ating textual summaries of physiological time series data to aid medical personnel in monitoring babies in neonatal intensive care units. Our studies suggest that sum- marization is a communicative task that requires data analysis techniques for de- termining the content of the summary. We describe a prototype system that summarizes physiological time series. 1 Introduction Time series data is ubiquitous — any measure- ment humans make over a period of time pro- duces a time series. We are building a system to summarize physiological times series data such as heart rate, and blood pressure measured in neonatal intensive care units. 2 Background The SumTimE project aims to develop generic techniques to produce textual summaries of time series data (Sripada et al, 2001). We initially worked in two domains, meteorology and gas turbines. In meteorology we generate textual weather forecasts from weather data such as wind speed, wind direction, and wave heights. In gas turbines we generate textual summaries of unex- pected patterns in sensor data such as exhaust temperature, liquid fuel flow, and turbine speed. In each of these domains we are working with industrial collaborators and have built prototype systems. Using the experience from both these domains we have now started working on physiological time series data in collaboration with NEONATE (Ewing et al, 2002) project. The main objective of NEONATE has been to produce a decision sup- port system for the medical personnel working in the neonatal intensive care unit (NICU). In the NEONATE project, a research nurse has been employed to collect data from the neonatal intensive care unit at Simpson Maternity Hospi- tal, Edinburgh using a software tool, BabyWatch (Ewing et al, 2002). Physiological parameters such as heart rate, mean blood pressure and tem- perature are recorded at one-second frequency using various probes attached to the baby. In order to monitor the health of babies, medi- cal personnel (doctors and nurses) working in the neonatal unit are required to inspect such data continually. Currently they use visual displays of the data. Our system will generate textual sum- maries of these data as an aid to the medical per- sonnel. We believe that interpreting textual summaries is lot quicker and does not require much mathematical (statistical) knowledge when compared to interpreting graphical displays. 3 Knowledge Acquisition We have carried out a variety of knowledge ac- quisition (KA) activities using multiple tech- niques developed in the expert system community (Scott, Clayton, and Gibson, 1991) to understand how humans perform data summari- zation. 167 20.0 60.0 BM(RD) 27 Feb 96 03:00� 03 10� 03 20� 0333� 3343� 03:50� 04:00� 04:10� 04:20� 04:30� 04:40� 04:50� 05 00� 0510� 3520� 3530� 05:40� 05:50 Figure 1. Plot of mean blood pressure Figure 1 shows a time series plot of mean blood pressure sampled every second for three hours. Figure 2 …

Content truncated to 500 words