Hypoxic-ischemic (HI) brain injury is a primary result of perinatal asphyxia (deficiency of oxygen and excess of carbon dioxide in the blood) and a significant contributor to poor neurodevelopment and severe neurological deficits in newborns. The cardiovascular response to hypoxia, which is important to prevent the core organs from the hypoxic insult, is variable between individuals, and the variability is associated with brain injury following hypoxia. It is hypothesized that the autonomic nervous system (ANS), especially the sympathetic branch (SNS), plays a determinant role in this variability. Understanding the physiological response to an HI insult may help with understanding the determinants of the subsequent neurodevelopmental outcome and with identifying fetuses and neonates who are at risk of HI brain injury. Another significant issue in perinatal research and clinical practice is how to accurately detect hypoxia, especially during the prenatal period when the fetus is not directly accessible. The features associated with the physiological response may improve the diagnosis of perinatal hypoxia. This thesis aims to:
1) Explore whether the SNS activity determines the inter-individual variation of the cardiovascular response to hypoxia and the subsequent neurological outcome;
2) Develop an accurate automated hypoxia detection method.
Heart rate variability (HRV) signals are defined as the variation of inter-heartbeat intervals, and reflect the autonomic control of cardiovascular function. In this thesis, the time–frequency distribution (TFD) of HRV signals was used to measure the continuous autonomic response to hypoxia and to extract effective non-stationary features for the detection of perinatal hypoxia. A high resolution TFD is critical for the accuracy of autonomic measurement and feature extraction. Therefore, a class of HRV-adapted TFDs was proposed using modified lag-independent kernels. In these TFDs, a new parameter is defined as the minimal frequency difference among signal components, and superiorly removes the cross-terms and improves the TF resolution as compared with existing good-performing TFDs that commonly used in HRV analysis. The proposed TFD can also enhance the discriminating ability of TFD-based HRV features in hypoxia detection.
To investigate the continuous SNS response to hypoxia in a neonatal piglet model with controlled hypoxia, the spectral power of the low frequency (LF) component of HRV signals was used to quantify the sympathetic regulation during hypoxia. The instantaneous component power was calculated based on the instantaneous frequency (IF) which was estimated using a component linking approach. The power of the LF component increased early during the period of hypoxia and ii then dropped to baseline levels prior to the heart rate (HR) and blood pressure (BP) reaching their maxima. There was no association between changes in the LF power and features of the cardiovascular response (HR pattern and BP index) to hypoxia or neurological outcome. These results suggest that the SNS contributes to the initial cardiovascular response to hypoxia in a stereotypical fashion but is then unable to maintain this contribution to the maintenance of adequate BP and organ perfusion. Therefore, the autonomic control (especially the sympathetic component) is not the source of inter-individual variation in the cardiovascular response to acute hypoxia since it is not the predominant factor in mediating the cardiovascular response to hypoxia.
Nevertheless, HRV features do reflect the autonomic changes in response to hypoxia, and this provides useful information for the detection of perinatal hypoxia. Features useful for hypoxia detection including those based on the IF and instantaneous amplitude of HRV signal components as well as those derived from matrix decomposition of the signals’ time–frequency distributions using singular value decomposition and non-negative matrix factorization were explored. An automated detection approach was developed by feeding the selected features into a support vector machine classifier. The proposed method was tested using an accurate dataset recorded from a neonatal piglet model with accurately measured and controlled hypoxia and there was a strong detection performance with sensitivity 93.3%, specificity 98.3% and accuracy 95.8%. This method outperforms methods based on stationary features and those using non-stationary features.