The present thesis discusses advanced polysomnogram signal processing approaches to perform computer-assisted sleep analysis and disorders characterisation and differentiation. Sleep electrophysiology and biomedical signal processing are branded by time-consuming assessments, inter-subjects' and inter-raters' variability. Those challenges compromise the analysis, characterisation and differentiation in terms of time analysis, tolerance to subjects' out-of-norm patterns and raters' biased sleep stages scoring. The proposed models explored computer-aided neuronal-cardiac synchronisation, sleep onset staging and arousal detection; looking forward to improve regular and abnormal sleep characterisation. Foremost, the differentiation of control, insomnia and schizophrenia groups and individuals was pursued throughout computational models and actual clinical datasets. Four core studies elaborated the literature background, methods, models, experimental datasets, data analysis and insights to attain sleep, groups and individuals characterisation. These interrelated studies built up an initial preprocessing model, going through sleep onset, insomnia characterization to achieve subjects and groups differentiation. The first core study focused on the computer-assisted preprocessing of polysomnogram signals. Preprocessing addressed the removal of embedded cardiac-type pulses and muscular artefacts, accompanied by the suppression or attenuation of the additive white noise from cortical and ocular signals. The proposed model successfully performed artefact removal upon 80% of the 1200 analysed electroencephalographic (EEG) and electrooculographic (EOG) epochs and 100% concerning background noise suppression. The second core study concerned the synchronisation between the central and autonomic nervous system to characterise control, insomnia and schizophrenia cohorts. The calculation of computer-assisted correlation, coherence and coupling metrics measured the neuronal-cardiac functional interdependence by means of EEG and electrocardiographic (ECG) signals. Those approaches represented linear, nonlinear and statistical methods. The spectral synchrogram derived from Wavelet-based coherence outperformed the characterisation of control and schizophrenia groups by tracking the interdependence of EEG alpha band to Heart Rate Variability Low Frequency (HRV-LF) as driving feature. The third core study centred in sleep onset staging for a computer-aided estimation of the sleep onset stages, and the computer-aided differentiation of control and insomnia individuals. Firstly, the approach introduced adaptive processing and machine learning tools for the automated sleep staging. Secondly, the approach suggested novel characterisation metrics for the differentiation of control, insomnia and schizophrenia cohorts dealing with out-of-norm patterns and raters' biased sleep stages scoring. The introduced biosignal modelling generalised polysomnogram signals made of thousands of time points with only 10 reconstructive coefficients. The performance metrics of control and insomnia subjects' logistic classifier highly rated: sensitivity (87%), specificity (75%) and accuracy (81%). The fourth core study appointed the computer-aided arousal marking, performing fuzzy logic-based detection. The model identified spontaneous arousals, EEG with chin tension and limb movement-related arousals. The model demonstrated an average Arousal Index (ArI) in regular sleepers around 20, which agreed with expert scored assessments. Overall, this original research work provides novel contributions in biomedical signal processing and a supporting role in the sleep electrophysiology towards a more comprehensive knowledge of insomnia and schizophrenia disorders.
展开▼
