Investigating the infant respiratory control system using non-linear analysis

Philip Terrill (2009). Investigating the infant respiratory control system using non-linear analysis PhD Thesis, School of Information Technol and Elec Engineering, The University of Queensland.

Attached Files (Some files may be inaccessible until you login with your UQ eSpace credentials)
Name Description MIMEType Size Downloads
s40063805_phd_correctedthesis.pdf Corrected thesis application/pdf 2.40MB 0
s40063805_phd_listofcorrections.pdf application/pdf 384.30KB 0
Author Philip Terrill
Thesis Title Investigating the infant respiratory control system using non-linear analysis
School, Centre or Institute School of Information Technol and Elec Engineering
Institution The University of Queensland
Publication date 2009-06-01
Thesis type PhD Thesis
Supervisor A/Prof Stephen J. Wilson
A/Prof David M. Cooper
Total pages 295
Total colour pages 13
Total black and white pages 282
Subjects 08 Information and Computing Sciences
Abstract/Summary Over recent years, the burden of disease associated with paediatric sleep disorders has become increasingly apparent. Sleep medicine is a diagnostically intensive field, where the gold standard investigation is polysomnography. Polysomnography is demanding in terms of human and physical resources, and is limited to metropolitan tertiary hospitals. The development of diagnostic systems reliant on only limited physiological instrumentation such as breathing patterns, which do not require continuous expert supervision, is highly desirable. Recently, it has been demonstrated that human breathing patterns follow non-linear dynamics, and that these dynamics are likely to be chaotic. Furthermore, it has been shown that non-linear measures of breathing patterns vary between human sleep states and in states of disease. The aims of this thesis were thus to: • Develop a robust mathematical tool able to quantify non-linear patterns in infant breathing • Characterise breathing during infant active sleep (AS) and quiet sleep (QS) by applying this non-linear tool to quantify non-linear features of breathing patterns • Determine whether there are non-linear indices which allow AS to be accurately discriminated from QS • Investigate the use of a non-linear discriminator as the mathematical basis for an automated sleep staging tool • Investigate whether non-linear indices, and therefore subsequent discriminators, vary with infant maturation Recurrence plot analysis was investigated as the basis for a robust mathematical tool to quantify non-linear dynamics in infant breathing patterns. Using a dataset of AS and QS periods of inter-breath interval (IBI) breathing from 32 healthy infants, the false nearest neighbour’s (FNN) method was applied with results suggesting that an embedding dimension in the range of 6-8 is appropriate for adequate unfolding of the attractor in phase space. Three threshold selection criteria were investigated: (1) fixed radius embedding; (2) fixed recurrence embedding; and (3) distance matrix normalisation with fixed recurrence embedding. Only fixed recurrence embedding ensured that recurrence plot saturation was consistent across a dataset with variable attractor structure, and ensured that estimates of dynamic invariants were not confounded by attractor size. Polysomnograms obtained for 32 healthy infants from the Collaborative Home Infant Monitoring Evaluation (CHIME) dataset were sleep staged by an experienced sleep clinician. Six periods each of AS and QS were identified for each infant, and data from the abdomen channel of respiratory inductive plethysmography extracted, and converted to Inter-Breath Intervals (IBIs). For each period, a radius vs. recurrence plot was generated, and Recurrence Quantification Analysis (RQA) applied using fixed recurrence embedding. Attractor diameter and distance separating close neighbours’, measured using the variable Radius (RAD) were larger for AS periods. Invariant to the chosen embedding dimension, or fixed recurrence value, results were higher for AS than QS for variables estimating sensitivity to initial conditions (Determinism, Maximum Diagonal Line Length and Average Diagonal Line Length), and quantifications of laminar periods and tangential motion (Laminarity, Trapping Time, Maximum Vertical Line Length). The ability of each of the RQA variables to discriminate AS from QS was assessed using receiver operator curve (ROC) analysis, and calculating the area under curve (AUC). This was also applied to the conventional statistical variables: mean IBI (IBIAVG), IBI variance, co-efficient of variation (IBICV) and inter-quartile range (IBIIQR). The best RQA discriminator was RAD (AUC=0.98); and laminarity (AUC=0.95). The best statistical discriminator was IBIIQR (AUC=0.95). Scaling according to IBIAVG improved results: RAD/IBIAVG (AUC=0.99), while IBIIQR/IBIAVG (AUC=0.98). To validate results from the CHIME data and to investigate maturation, RQA analysis, and discrimination assessment were applied to data from the Mater Children’s Hospital Prospective Cohort Normal Sleep and Breathing Study (MNSBS) in which 24 infants had full overnight polysomnograms at 2wks (S1), 3mnths (S2) and 6mnths (S3). RQA confirmed the trends in the CHIME data, with differences between AS and QS less distinct at S1. The discrimination ability of variables was consistent with CHIME results and optimal discrimination achieved at S2. The best discriminator was RAD/IBIAVG (AUC=0.979, 0.999, 0.983 for S1, S2 and S3), outperforming the best statistical discriminator IBIIQR/IBIAVG (AUC=0.965, 0.990, 0.959 for S1, S2 and S3). The sleep state discriminators based on recurrence plot analysis variables were assessed for potential as the mathematical basis for automated sleep state classification system- firstly in the CHIME dataset using a simple threshold based classifier to classify known periods of sleep as active sleep and quiet sleep; and secondly in the MNSBS using linear discriminant analysis to develop a multivariate classifier to classify 30 sec epochs of overnight PSGs as wake, active sleep and quiet sleep. In the CHIME dataset, a RAD/IBIAVG based classifier could stage periods of known sleep as AS/QS with an average agreement of 88% (kappa=0.75). In the MNSBS dataset, a multivariate combination of the variables was able to classify Wake/AS/QS with 77.5% (κ=0.65), 87.5% (κ=0.79) and 84.8% (κ=0.75) agreement with the human expert scoring at 2 weeks, 3 months and 6 months respectively. It performed considerably better than a comparison classifier based on IBI statistics and spectral density features proposed in the literature, and makes clinically significant improvements over respiratory only sleep state classifiers previously proposed in literature. The non-linear sleep state discriminators identified in this thesis enable the automatic classification of infant sleep states with greater accuracy than conventional statistical and frequency spectrum quantifications. It is concluded that this automated sleep staging tool has practical applicability to a minimal channel diagnostic screening tool, as an adjunct to reduce workload in conventional PSG, and in population based sleep research. Its requirement for only a single channel of un-calibrated breathing data makes it ideal for studies outside a dedicated sleep laboratory.
Keyword paediatric sleep
respiratory control
breathing control
Non-linear analysis
recurrence plot
automated sleep state classification
Additional Notes The following pages (according to the pdf page number, not the labelled number in the document) should be printed in colour: 30, 33, 34, 36, 39, 42, 115, 137, 139, 140, 218, 220, 221

Citation counts: Google Scholar Search Google Scholar
Created: Wed, 08 Dec 2010, 03:05:37 EST by Mr Philip Terrill on behalf of Library - Information Access Service