Use of anesthesia data for research

Abstract

In CHAPTER 2 we have developed reference curves for blood pressure during anesthesia in children. We collected pediatric anesthesia data from 10 different hospitals in two different countries. With 116,362 anesthetic procedures we constructed sex specific references, adjusted for age, weight or height. The curves provided in CHAPTER 2, help clinicians with a better reference for intra-operative blood pressure monitoring compared to what was previously available. These references can also be used in research, as we illustrated in CHAPTER 3. We applied the references to a single-center pediatric cohort, to explore which children get a relatively low blood pressure during anesthesia. The blood pressure was adjusted for sex and height using the references from CHAPTER 2. We studied the association of patient and anesthesia procedure characteristics with this normalized blood pressure. Part of the variation in blood pressure could be explained by the procedure characteristics, such as application of a loco-regional technique. Only small effects of patient characteristics on blood pressure were estimated, therefore we could not define a typical child that develops low blood pressure during anesthesia. In PART II of this dissertation we focused on artifacts in physiologic measurements (i.e. heart rate, saturation, end tidal carbon dioxide, non-invasive blood pressure and invasive blood pressure) during anesthesia. First, in CHAPTER 4 we observed the incidence of artifacts in pediatric cases during 170 hours of anesthesia. Incidence of artifacts ranges from 0.5% for heart rate to 7.5% for end tidal carbon dioxide. We found that these artifacts did not occur at random and were dependent on different factors, e.g. type of measurement, age of the patient, type of surgery or the phase of surgery at which the measurement was done. As some of these factors could also be related to outcome, we hypothesized that artifacts or the way artifacts are filtered, could act as a confounder. Second, in CHAPTER 5 we tested this hypothesis by studying the effect of different artifact filtering methods on the result of an example study. The artifact filtering methods were identified with a systematic literature search. As an example we used hospital data from adults older than 60 who underwent medium to high risk surgery and analyzed the relation between hypotension and postoperative myocardial injury. We showed that there is indeed a small systematic effect of artifact filtering methods on the estimated relation between hypotension and myocardial injury. Finally, in CHAPTER 6 we focused in more detail on the artifacts in invasive blood pressure data and formulated a method to identify artifacts automatically. We showed that besides dependency on factors (e.g. type of surgery and phase of surgery) it also matters when and how artifacts are identified. When someone observes the procedure live, similar to CHAPTER 4, the artifact identification was different than when someone reviewed the data retrospectively, which is customary in database research. The information or context available live and retrospective is different, and therefore the conclusion drawn by an observer is different. Therefore we cannot simply use live or retrospective manual annotations as a golden standard for artifact identification. A clear definition of what is assumed to be an artifact should first be formulated and reported by a researcher. Despite the differences we hypothesized that the process of artifact identification could still be automated. In CHAPTER 6 we applied different learning algorithms to model the identification of artifacts. In this study, the performance of these algorithms remained mediocre at best. Future research could focus on development of better performing algorithms using additional information and further optimization of the training of such algorithms to reduce the manual work required.