Sleep stages classification using deep learning on wearable sensors data in patients with sleep disorders

Sleep plays a fundamental role in physical and mental health, and poor-quality sleep is associated with a wide range of chronic conditions. However, sleep disorders remain underdiagnosed and undertreated in many cases, in part due to the limited accessibility of accurate sleep monitoring. Although Polysomnography (PSG) is the gold standard for sleep monitoring, it is limited to controlled clinical settings, requires trained professionals, and is typically restricted to a single night of sleep that may not reflect the typical sleep behavior of the patient. Recent advances in wearable technology offer new opportunities for continuous, unobtrusive, and home-based sleep monitoring. This thesis explores a deep learning-based approach to automatic sleep staging using physiological signals collected from wearable devices to provide accessible tools for sleep assessment, particularly for individuals with suspected or diagnosed sleep disorders. The proposed method uses a convolutional neural network architecture, the U-Sleep, trained on multimodal data - Acceleration, Blood Volume Pulse, Electrodermal Activity, and Skin Temperature - recorded using the Empatica E4 wristband. The dataset includes 127 participants with simultaneously recorded PSG and wearable data, allowing for direct performance comparison with the clinical standard. Results show that the model can accurately estimate key sleep parameters, with Bland-Altman analysis revealing good agreement for Sleep Efficiency, REM Latency, Wake After Sleep Onset, and durations of REM and Deep sleep. Epoch-by-epoch concordance reached an accuracy of 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 the whole dataset, and 0.77±0.05 and 0.74±0.06 for healthy participants, respectively. To assess the robustness and generalizability of the model, additional experiments were conducted on external datasets and with a different wearable device. These evaluations confirmed the adaptability of the model to various data sources and configurations, highlighting its potential for scalable application in real-world contexts. In conclusion, this work demonstrates the feasibility of using multimodal wearable data for personalized sleep staging. By capturing individual-specific physiological patterns, the proposed approach supports the use of deep learning for precision sleep health, with potential applications in long-term monitoring, early detection, and management of sleep disorders in a clinical setting.