Movement characterisation with accelerometry for improved paediatric sleep assessment

Lamprecht, Marnie L. (2016). Movement characterisation with accelerometry for improved paediatric sleep assessment PhD Thesis, School of Information Technology and Electrical Engineering, The University of Queensland. doi:10.14264/uql.2016.783

Attached Files (Some files may be inaccessible until you login with your UQ eSpace credentials)
Name Description MIMEType Size Downloads
s4119125_final_thesis.pdf Thesis (open access) application/pdf 28.31MB 0
Author Lamprecht, Marnie L.
Thesis Title Movement characterisation with accelerometry for improved paediatric sleep assessment
School, Centre or Institute School of Information Technology and Electrical Engineering
Institution The University of Queensland
DOI 10.14264/uql.2016.783
Publication date 2016-09-26
Thesis type PhD Thesis
Supervisor Philip I. Terrill
Andrew P. Bradley
Total pages 224
Language eng
Subjects 0906 Electrical and Electronic Engineering
0903 Biomedical Engineering
Formatted abstract
Actigraphy is increasingly used to non-invasively estimate sleep quality in children with a suspected sleep disorder. Commercial actigraphs summarise wrist movement, conventionally measured with a uni-axial accelerometer, within a fixed epoch (typically 30s). Wake is subsequently identified as epochs of increased activity and sleep is identified as epochs of inactivity. This classification framework has some distinct limitations: actigraphy misclassifies activity during sleep as wake, and inactivity during wake (i.e. quiet rest) as sleep. In this thesis we will address these limitations by investigating three hypotheses. Firstly, uni-axial accelerometry measured solely at the wrist restricts prediction accuracy, since movements orthogonal to the measurement axis, or occurring elsewhere on the body, cannot be detected. Utilising multisite tri-axial accelerometry may consequently improve sleep and wake prediction. Secondly, there are movement characteristics that can differentiate sleep from wake because the physiological nature of these movements differ. Identifying these characteristics may reduce false wake detections. Finally, physiological and pathological events such as apnoeas, hypopneas and transient arousals may be associated with sleep movements that contribute to false wake detections. Exploring this association may consequently explain the presence of some sleep movements.

In order to address the hypotheses, 38 participants (27 male, aged 5 − 16 years) were recruited from children attending the sleep laboratory for suspected sleep-disordered breathing. These children were studied concurrently with polysomnography and a custom system (synchronised to within 0.1s) that recorded raw tri-axial accelerometry data (8−bit, 100Hz, ±2G) simultaneously at the left index fingertip, left wrist, upper thorax, left ankle and left great toe.

The first analysis compared the accuracy of predicting sleep and wake epochs with uni- axial, tri-axial, and multisite accelerometry. Tri-axial versions of the conventional 30s epoch summaries were derived and compared to conventional uni-axial accelerometry. Multisite data were explored and verified using two feature selection algorithms with the tri-axial summaries for each accelerometer. Classification performance was significantly improved when incorporating additional accelerometers, and measuring movement with tri-axial accelerometry (Kappa agreement for multisite, tri-axial and uni-axial accelerometry: 0.565(0.231), 0.402(0.141) and 0.268(0.210), p < 0.05). Tri-axial accelerometry has clear benefits with no increase in cost or invasiveness. Although multisite accelerometry provides additional performance benefits, these benefits come at the expense of system complexity and patient discomfort.

Moving away from epoch-by-epoch predictions, the second analysis assessed wake detection on a movement-by-movement basis. Localised spectral characteristics of raw segmented wrist movements were identified using the discrete wavelet transform. Characteristics that significantly differed between sleep and wake movements were then used to predict wake on a movement-by-movement basis. In general, short-duration wake movements had regions of increased spectral energy, were more vigorous, and consistently had spectral content characteristic of limb positional changes. However, predicting wake on a movement-by-movement basis had similar performance to the 30s activity counts (area under the receiver operating characteristics curve: 63.9(6.7)% vs. 69.7(7.9)% respectively). The similar performance of these distinctly different approaches, together with the consistently average predictive performance seen throughout the analyses, shows that movement information cannot accurately estimate sleep in a generalised classification model.

The final analysis explored possible causes of the confounding sleep movements by analysing the temporal association with transient arousals, apnoeas and hypopneas manually scored from polysomnography. On average, 21.4% of apnoeas, 40.8% of hypopneas and 67.5% of arousals coincided with wrist movement. However, the prevalence and corresponding associations varied considerably across the cohort. Arousals during sleep that were associated with movement were generally longer than other arousals (12.2s vs. 7.9s, p < 0.01). Similarly, movements that occurred during an arousal were longer than other sleep movements (9.56s vs. 2.35s, p < 0.01). The association between lengthy arousals and lengthy sleep movements suggests that these longer arousals contribute to false wake detections. Although actigraphy cannot predict all arousals, it can likely predict the lengthier arousals that disrupt sleep.

We can conclude from the analyses in this thesis that multisite tri-axial accelerometry offers distinct performance benefits for sleep assessment; however, the associated practical compromise from additional accelerometers may not be appealing for a device targeted at home-based sleep assessment. Transient arousals are strongly associated with the lengthier sleep movements that confound sleep estimates with commercial actigraphy. Considering that these arousals are characteristic of sleep disturbance, and actigraphy likely detects these events, incorporating the detection of these longer sleep movements into the actigraphy scoring routine may capture the severity of sleep disturbance associated with a sleep disorder. Existing actigraphy systems only estimate sleep quality; however, future actigraphy systems will likely benefit from identifying signs indicative of sleep disorder severity.
Keyword Actigraphy
Sleep disturbance
Sleep fragmentation
Automated sleep scoring
Time-series signal analysis
Wavelet analysis
Sleep disorder detection

Document type: Thesis
Collections: UQ Theses (RHD) - Official
UQ Theses (RHD) - Open Access
Version Filter Type
Citation counts: Google Scholar Search Google Scholar
Created: Thu, 15 Sep 2016, 21:12:33 EST by Marnie Lamprecht on behalf of Learning and Research Services (UQ Library)