Florey Department of Neuroscience and Mental Health - Theses
Permanent URI for this collection
Search Results
1 - 1 of 1
-
Item‘Somnivore’ a user-friendly platform for automated scoring and analysis of polysomnography data( 2016)The low-throughput nature of manual scoring of polysomnography (sleep) data, both in terms of speed and consistency, is a major factor preventing sleep research from reaching its full efficiency and potential. Automated approaches developed previously have generally failed to provide sufficient accuracy or 'usability' for sleep scientists lacking specialist-engineering expertise. Moreover, all earlier approaches have only been validated using baseline data, suggesting a failure to embed in the algorithm the robustness to remain effective when used to analyse the effect on sleep of treatment or disease. Finally, no single approach has been validated for mouse, rat and human data. Therefore, the aim of my research was to develop a user-friendly platform for real-time automated scoring and analysis of polysomnography data. The program is known as ‘Somnivore’ (from Latin somnus, ‘sleep’, and vorare, ‘to devour’), and was developed using state of the art supervised machine learning technology, with support vector machine (SVM) at its core, and coded as a graphical user interface (GUI)-based solution in the Matlab™ ambient. Somnivore learns, in parallel, by surveying features from a variety of different inputs (including EEG, EMG, EOG and ECG) and outputs data into the various sleep stages (wake, NREM, N1, N2, N3, REM). The classifier is trained for each subject via a brief session of manual scoring. Design and development strategies were built around both theoretical and heuristic approaches. This led to a multi-layered system capable of learning from extremely limited training sets, using large input space dimensionalities from a rich variety of polysomnography inputs, and with rapid computational times. Validation was pursued to approach the numerous contentious dynamics that have led to the demise of previous solutions. Somnivore generalisation was evaluated at the level of canonical classifier evaluation metrics such as F-measure, as well as experimental end-measures more germane to traditional biological sleep research. Somnivore, generated superior generalisation, with high power, on both murine (n = 54) and human (n = 52) recordings. These included multiple rat strains (Sprague-Dawley, Wistar) and mouse phenotypes (wild type, orexin neuron-ablated transgenic), various pharmacological interventions (placebo, alcohol, muscimol, caffeine, zolpidem, almorexant), and in humans, both genders, younger and older subjects, and subjects with mild cognitive impairment (MCI). Somnivore’s generalisation was also evaluated in conditions of signal challenged data, and provided excellent performance in all conditions using only one EEG channel for learning. Remarkable results were also reported for learning undertaken using only one EMG channel or two EOG channels. Furthermore, validation studies highlighted that a substantial part of the disagreement between manual and automated hypnograms was located within transition epochs. As Somnivore has several features geared towards the management of transition epochs, further control over generalisation is also possible. Comprehensive inter-scorer agreement analysis was conducted on human data, showcasing how inter-scorer agreement between manual hypnograms and their automated counterparts provided by Somnivore is comparable to the gold-standard of the inter-scorer agreement between two experts trained in the same laboratory. Results also highlighted critical problems within the scoring of stage N1. However, inter-scorer agreement validation studies also confirmed what has already been reported in the literature, that N1 is a volatile stage that systematically produces inadequate agreement even between trained experts, both within or outside the same laboratory. Accordingly, Somnivore performed as well on N1 as reported in the literature for manually scored data. Due to the high-throughput nature of Somnivore’s analyses of experimental end-measures, several novel, cautionary findings were extracted from the recordings provided by external laboratories for this research evaluations. Additionally, as Somnivore is also capable of scoring real-time during polysomnography recordings, it will facilitate the development of more advanced protocols such as biofeedback sleep-deprivation protocols and integrated optogenetics. In conclusion, Somnivore, has been comprehensively validated as an accurate, reliable, high-throughput solution for scoring and analysis of polysomnography data, in a range of experimental situations including studies of normal physiology and tests related to drug discovery for the improved treatment of sleep disorders and psychiatric diseases.