Sleep Stage Classification from Wristband Sensor Data in Patients with Sleep Disorders

Oriella Gnarra, Camilla Massaro, Jan D. Warncke, Tobias Nef, Markus Schmidt, Diego Paez-Granados

Confirmation: I have read and agree with the IEEE BHI 2025 conference submission's policy on behalf of myself and my co-authors.

Keywords: Data Analysis, Deep Learning, Sleep Stages, Sleep Disorders, Wearables.

Abstract: Polysomnography (PSG) is the clinical gold standard for sleep assessment, but it is constrained to single-night recordings in hospital environments and requires trained personnel. The emergence of wearable technologies enables unobtrusive and potentially long-term sleep monitoring. This study presents an AI-based approach tailored to individuals with sleep disorders, evaluating the U-Sleep neural network trained on multimodal signals: Acceleration (ACC), Blood Volume Pulse (BVP), Electrodermal Activity (EDA), and skin Temperature (TEMP), collected via the Empatica E4 wristband. The goal is to advance precision health by developing sleep monitoring tools adaptable to individual physiological profiles, particularly in patients with suspected or diagnosed sleep disorders. This is the first study of multimodal sleep stage classification combining EDA and TEMP with ACC and BVP, which was developed using data from 127 participants undergoing simultaneous PSG and wearable recordings. Evaluating and verifying an improved performance through the integration of diverse physiological signals in a personalized model for individuals with sleep disorders. Bland-Altman analysis revealed an overestimation of Light sleep and an underestimation of Sleep Onset Latency (SOL). However, sleep efficiency, REM latency, wake after sleep onset, REM, and Deep sleep durations were estimated without significant differences from PSG. Epoch-by-epoch accuracy reached 0.87±0.07 for Wake, 0.90±0.04 for REM, 0.71±0.07 for Light, and 0.89±0.04 for Deep sleep. Overall accuracy and F1-score were 0.69±0.08 and 0.62±0.11 for patients and 0.77±0.05 and 0.74±0.06 for healthy participants, respectively. These results highlight the feasibility of applying multimodal wearable data for accurate sleep staging across diverse populations. By enabling models to capture individual-specific patterns, this work contributes to the field of precision sleep health and lays the groundwork for remote, patient-tailored management of sleep disorders. Although further refinement is needed before clinical application, the approach shows promise for advancing AI-driven, personalized diagnostics in real-world settings.

Track: 1. Biomedical Sensor Informatics

Registration Id: X3NS3K5TZTM

Submission Number: 278