Unconstrained Sleep Staging Based on Contextual Scalograms Using Piezoelectric Film Sensors

Sleep staging plays a critical role in evaluating sleep quality and in the prevention and diagnosis of sleep-related disorders. In this study, we recruited healthy young adults and collected a total of 62 nights of sleep data using a sleep monitoring pad embedded with piezoelectric film sensors (PEFS) alongside a standard multi-channel polysomnography (PSG) system. Contextual scalograms were generated from the PEFS signals via continuous wavelet transform (CWT). To capture temporal dependencies within the time-frequency representations, we integrated a Long Short-Term Memory (LSTM) module into the GoogLeNet architecture. The proposed Go-LSTM network achieved an average classification accuracy of 71.12% across four sleep stages, with a Cohen's kappa coefficient of 0.53. These results suggest that the Go-LSTM model based on PEFS signals has promising potential for automatic sleep stage classification and may enable long-term, non-intrusive sleep monitoring in home settings.