Pregnancy-related complications affect approximately 15% of pregnancies and, if severe, can have long-term consequences. Timely recognition of physiological deterioration is known to reduce the prevalence and severity of complications. However pregnancy-associated changes in vital signs (blood pressure, heart rate, temperature, oxygen saturation, and respiratory rate) complicate the detection of abnormal physiology, and these changes are not well documented. This thesis describes the development of algorithms to ensure the collection of good-quality vital-sign data during the antenatal and postnatal stages of pregnancy, and the design of an evidence-based obstetric early warning score. Vital-sign information from 1,000 pregnant women during pregnancy, labour, and after delivery was collected during the 4P study using pulse oximetry, oscillometry for blood pressure measurement and a tympanic thermometer. Dynamic time warping was used to assess beat-by-beat quality in the photoplethysmograph (PPG) waveform obtained from the pulse oximeter. The resulting signal quality index enabled the exclusion of poor-quality sections and their associated measurements of heart rate and peripheral oxygen saturation. A robust measurement of respiratory rate was obtained by combining information from the PPG waveform, and accelerometer and gyroscope waveforms from a smartphone. After processing, frequency-based techniques, such as Fourier analysis and auto-regressive modelling, and time-domain peak detection were fused to estimate respiratory rate. When compared with the reference respiratory rate obtained from midwife measurement, the lowest mean absolute error of 1.16 breaths per minute was obtained from respiratory rate estimates from the y-axis of the gyroscope. Antenatal and postnatal reference ranges for each vital sign were developed with a standard polynomial multilevel (hierarchical) model using 10,000 vital sign measurements from 620 healthy women in the 4P study. Vital-sign trajectories confirmed known trends of blood pressure and heart rate changes during pregnancy, and provided new information about other vital-sign trends. Additional covariates were included to investigate the effect of parity and body mass index (BMI) on vital-sign trends. The outer centiles of the vital sign reference ranges were used to design an obstetric early warning score (C-ObsEWS) that took into account gestational age or time after delivery. The investigations in this thesis contribute additional knowledge of pregnancy-associated vital-sign changes, and lead to an initial proposal for an evidence-based obstetric early warning score specific to the stage of pregnancy.

Cerebral autoregulation describes the clinically observed phenomenon that cerebral blood flow remains relatively constant in healthy human subjects despite large systemic changes in blood pressure, dissolved blood gas concentrations, heart rate and other systemic variables. Cerebral autoregulation is known to be impaired post ischaemic stroke, after severe head injury, in patients suffering from autonomic dysfunction and under the action of various drugs. Cerebral auto-regulation is a dynamic, multivariate phenomenon. Sensitive techniques are required to monitor cerebral auto-regulation in a clinical setting. This thesis presents 4 related signal processing studies of cerebral autoregulation. The first study shows how consideration of changes in blood gas concentrations simultaneously with changes in blood pressure can improve the accuracy of an existing frequency domain technique for monitoring cerebral autoregulation from spontaneous fluctuations in blood pressure and a transcranial doppler measure of cerebral blood flow velocity. The second study shows how the continuous wavelet transform can be used to investigate coupling between blood pressure and near infrared spectroscopy measures of cerebral haemodynamics in patients with autonomic failure. This introduces time information into the frequency based assessment, however neglects the contribution of blood gas concentrations. The third study shows how this limitation can be resolved by introducing a new time-varying multivariate system identification algorithm based around the dual tree undecimated wavelet transform. All frequency and time-frequency domain methods of monitoring cerebral autoregulation assume linear coupling between the variables under consideration. The fourth study therefore considers nonlinear techniques of monitoring cerebral autoregulation, and illustrates some of the difficulties inherent in this form of analysis. The general approach taken in this thesis is to formulate a simple system model; usually in the form of an ODE or a stochastic process. The form of the model is adapted to encapsulate a hypothesis about features of cerebral autoregulation, particularly those features that may be difficult to recover using existing methods of analysis. The performance of the proposed method of analysis is then evaluated under these conditions. After this testing, the techniques are then applied to data provided by the Laboratory of Human Cerebrovascular Physiology in Alberta, Canada, and the National Hospital for Neurology and Neurosurgery in London, UK.