Thesis on data visualisation and HRV analysis

[Raymond Ben]

The following PhD thesis is available from
http://www.eleceng.adelaide.edu.au/Personal/braymond/thesis.pdf.gz
It deals mostly with the application of data visualisation techniques to a
biomedical problem (heart rate variability analysis) but does delve a little
into issues of training and regularisation in the LSS and GTM.
Comments welcome.
Ben 


Abstract

Variations in heart rate on a beat-to-beat basis reflect variations in
autonomic tone. Heart rate variability (HRV) analysis is a popular research
tool for probing autonomic function and has found a wide range of
applications. This thesis studies data visualisation and classification of
HRV as alternatives to established processing methods.
Data visualisation algorithms transform a high-dimension data set into an
easily-visualised, two-dimensional representation, or mapping. This
transformation is conducted such that the interesting structure of the data
set is preserved. Visualisaton techniques thus allow the researcher to
investigate relationships between HRV data without the need to define fixed
bands of interest within each spectrum. Two visualisation algorithms are
primarily used throughout this thesis: the least-squares scaling (a form of
multidimensional scaling) and the generative topographic mapping (GTM).
The least-squares scaling (LSS) may be implemented using radial basis
function neural networks, adding the ability to project new data onto an
existing mapping. The training of such networks can be done in conjunction
with the construction of the map itself, or as a separate step. It has
previously been suggested that the former approach yields smoother networks
and thus better generalisation; here, it is shown that, with appropriate
network regularisation, the generalisation properties of the two methods are
comparable. It is also shown that the incorporation of prior knowledge (such
as class labels) into the LSS can improve the visual properties of the
resulting mapping. A simple modification to the GTM is given to allow
similar use of prior information.
The visualisation of HRV data is demonstrated on two data sets. The first
was from a simple study involving postural and pharmacological intervention
in healthy subjects. The LSS and GTM both produced logical mappings of the
data, with the ordering of the points within the map reflecting the
sympathovagal balance during the various phases of the study. Data
visualisation is also demonstrated on HRV data from overnight studies into
the sleep apnoea/hypopnoea syndrome. The ordering of the points within the
map in this case was strongly related to the power in the very low frequency
region of the spectra, known to be an indicator of sleep apnoea. Subjects
who suffered predominantly hypopnoeas rather than true apnoeas were found to
show HRV similar to control subjects.
The final section of the thesis briefly addresses the classification of HRV,
emphasising the combination of HRV with information from other diagnostic
signals and sources. Classification of data from the intervention study
showed that mean heart rate together with HRV allowed more reliable
classification than did either mean heart rate or HRV alone. In the
classification of sleep apnoea data, the addition of body mass index and age
did not improve classification; however, the inclusion of oxyhaemoglobin
desaturation information did improve the classification accuracy.