Analysing the efficacy of an advanced patient monitoring system to safeguard patient and operation outcomes

Abstract

Hazards and complications that can arise during anaesthesia are likely to cause adverse e ects and can lead to poor health outcomes. Advances in computing and recent research means these adverse reactions can be detected using expert systems. These systems alert clinicians of potential risks and developments, aiding the clinician's own judgement. This study presents a method for analysing the potential bene ts of expert systems using recorded operation data before they are used in a clinical environment. A data pipeline was developed for measuring the e cacy of these expert systems, in this case, the expert system Early Detection and Decision Information (EDDI) was evaluated. The development of the pipeline uses a database of 80,000 surgical procedures provided by Auckland District Health Board (ADHB). Using the pipeline, the physiological data provided in the database is run through the alert system within EDDI. The results from retrospectively applying EDDI to the physiological data is then used to analyse operations and investigate the relationship between the drugs administered and the alerts from EDDI. This research uses a confusion matrix and the corresponding Matthews correlation coe cient to quantify these relationships. This analysis approach can be applied to the expert system has a whole (i.e., to all the alerts provided at once) or for each alert individually. For the analysis of both the overall system and individual alerts, it is assumed that the expert system EDDI provides similar information (for the alerts it provides) to the clinical expertise of the anaesthetist. Here an investigation looks at if a closer \alignment" between clinician actions and EDDI relates to a better outcome for the operation. In this thesis the outcome of interest is operation duration and analysis showed that there were some interactions between \alignment" and operation duration, but the nature of this relationship could not be determined. The pipeline shows potential for further analysis of expert systems, other suggested clinical outcomes worth exploring using the approach presented here include hospital length of stay and 90-day postoperative mortality.