Un modello di apprendimento profondo end-to-end per la classificazione automatica degli stadi del sonno utilizzando segnali fisiologici minimi = An End-to-End Deep Learning Model for Automatic Sleep Stage Classification Using Minimal Physiological Signals

Sleep plays a key role in overall health and well-being, which makes it a significant topic for academic investigation. Polysomnography (PSG) is commonly used by researchers to analyse sleep, as it offers detailed insights via multiple physiological signals. However, PSG is expensive, technically complex, and typically confined to clinical environments for a single night. To enhance accessibility and enable long-term monitoring, data from commercially available wearable sensors can be utilized. Yet, for these sensors to reliably detect and interpret sleep patterns or diagnose sleep disorders robust algorithms are required. Such algorithms must accurately analyse sensor-derived signals to classify or predict various sleep stages. This study proposes an end-to-end deep learning model for automatic sleep stage classification using minimal physiological signals. Publicly available datasets comprising data from accelerometers and photoplethysmography (PPG) sensors were employed to facilitate broader real-world applicability across both healthy and clinical populations. The key contribution is learning data representations directly from raw signals through neural networks, rather than relying on manually engineered features. This approach is particularly advantageous when domain knowledge is limited, as convolutional neural networks can automatically discover robust patterns from raw data without requiring specialized feature-design expertise. Focusing on heart rate and body movements, our model processes 30-second epochs to classify wake, combined N1, N2, N3, and REM stages. Performance was validated on a hold-out dataset by comparing the model’s predictions to technician-annotated PSG recordings, sampled every 30 seconds. The results indicate that consumer-grade technology, when combined with carefully designed deep learning approaches, can complement, or potentially replace PSG for diagnosing and monitoring sleep disorders over multiple nights.