Sleep onset detection based on Time-Varying Autoregressive models with particle filter estimation

In this paper, we introduce a computer-assisted approach for the characterisation of sleep onset periods. The processing of polysomnographic (PSG) recordings involves the modelling of Time-Varying Autoregressive Moving Average (TVARMA) processes with recursive particle filtering. The feature set engages the computation of electroencephalogram (EEG) frequency bands δ, θ, α, ς, β, mean amplitude of electrooculogram (EOG) and electromyogram (EMG) signals. This is subsequently transferred to an ensemble classifier to detect Wake (W), non-REM1 (N1) and non-REM2 (N2) sleep stages. As a result, novel contributions in terms of non-Gaussian modelling of biosignal processes, approximation to PSG distributions with particle filtering and time-frequency analysis by complex Morlet wavelets within sleep staging, are discussed. The findings revealed performance metrics achieving in the best cases 93:18% accuracy, 6:82% error and 100% sensitivity/specificity rates.