Objective: To quantify and differentiate control and insomnia sleep onset patterns through biomedical signal processing of overnight polysomnograms. Methods: The approach consisted of three tandem modules: 1) biosignal processing module, which used state-space time-varying autoregressive moving average (TVARMA) processes with recursive particle filter; 2) hypnogram generation module that implemented a fuzzy inference system (FIS); and 3) insomnia characterisation module that discriminated between control and subjects with insomnia using a logistic regression model trained with a set of similarity measures (d1, d2, d3, d4). The study employed sleep onset periods from 16 control and 16 subjects with insomnia. Results: State-spaced TVARMA processes with recursive particle filtering provided resilience to nonlinear, nonstationary and non-Gaussian conditions of biosignals. FIS managed automated sleep scoring robust to inter-subjects' and inter-raters' variability. The similarity distances quantified in a scalar measure the transitions amongst sleep onset stages, computed from expert and automated hypnograms. A statistical set of unpaired two-tailed t-tests suggested that distances d1, d2 and d3 had larger statistical significance (pd1 < 6.5 × 10-5 , pd2 < 2.1 × 10-4 , pd3 < 4.5 × 10-7 ) to characterise sleeping patterns. The logistic regression model classified control and subjects with insomnia with sensitivity 87%, specificity 75% and accuracy 81%. Conclusion: Our approach can perform a supportive role in either biosignal processing, sleep staging, insomnia characterisation or all the previous, coping with time-consuming procedures and massive data volumes of standard protocols. Significance: The introduction of graph spectral theory and logistic regression for the diagnosis of insomnia represents a novel concept, attempting to describe complex sleep dynamics throughout transitions networks and scalar measures.
↧