Evaluation and extension of an automated system to diagnose the health of premature babies
Premature babies are not fully equipped to deal with the outside world. Their immature bodies make them highly at risk to a large number of physiological complications. Therefore, while they are still immature, premature babies are taken care of in neonatology intensive care units, where they are kept in a controlled and protective environment. In these units, the babies' vital parameters, such as their heart rate and temperature, are tightly monitored. The devices used to record this data are fitted with alarms that are triggered automatically when something wrong is detected. However, the alarm systems currently used are often unable to discriminate between real physiological problems and artefacts that should be ignored (e.g. when a measurement device drops-out). As a result, these systems produce a high rate of false alarms, where the alarm is triggered when nothing is clinically wrong. The can have unwanted consequences: first, it creates a noisy environment, which may disturb the babies; second, it can lead to the alarm being ignored when there is actually something wrong. In order to address this problem, my host laboratory has constructed a system that can automatically infer the state of health of premature babies in neonatology units, based on their recorded physiological data. The new system is designed to be able to deal with artefactual changes in the measurements, hopefully resulting in a reduced number of false alarms. In this thesis, I describe my contribution to this project, which is twofold. The first part deals with the problem of evaluating the performance of the system. To do this, I developed an online feedback process, where clinicians could provide information about the true clinical interpretation of the physiological data, which could be compared to the output of the system. The feedbacks provided by the clinicians were then analysed to investigate how the system could be improved. The second part of my thesis was motivated be early feedbacks from the clinicians, which indicated that the system was unable to deal with changes in the humidity measurements that occurred when the incubator door was opened and closed. To address this, I extended the current system, so that it was able to account for these changes, thereby increasing its capability to detect "true" physiological problems during the period after the incubator door is closed.
Laboratory of computational neuroscience, EPFL. - Carried out in the Institute for Adaptive and Neural Computation, at Edinburgh University's School of Informatics, under the supervision of Chris Williams. - External expert: Chris Williams
Record created on 2012-01-04, modified on 2016-08-09