High Frequency Physiological Data Quality Modelling in the Intensive Care Unit

Abstract

Intensive care medicine is a resource intense environment in which technical and clinical decision making relies on rapidly assimilating a huge amount of categorical and timeseries physiologic data. These signals are being presented at variable frequencies and of variable quality. Intensive care clinicians rely on high frequency measurements of the patient's physiologic state to assess critical illness and the response to therapies. Physiological waveforms have the potential to reveal details about the patient state in very fine resolution, and can assist, augment, or even automate decision making in intensive care. However, these high frequency time-series physiologic signals pose many challenges for modelling. These signals contain noise, artefacts, and systematic timing errors, all of which can impact the quality and accuracy of models being developed and the reproducibility of results. In this context, the central theme of this thesis is to model the process of data collection in an intensive care environment from a statistical, metrological, and biosignals engineering perspective with the aim of identifying, quantifying, and, where possible, correcting errors introduced by the data collection systems. Three different aspects of physiological measurement were explored in detail, namely measurement of blood oxygenation, measurement of blood pressure, and measurement of time. A literature review of sources of errors and uncertainty in timing systems used in intensive care units was undertaken. A signal alignment algorithm was developed and applied to approximately 34,000 patient-hours of simultaneously collected electroencephalography and physiological waveforms collected at the bedside using two different medical devices.