PERSONALIZED SLEEP SIGNATURE: A NOVEL APPROACH TO UNVEIL PAEDIATRIC SLEEP BEHAVIOUR WITH TRANSFORMER ATTENTION MECHANISM AND GRAPH ATTENTION NETWORKS

Sleep disorders manifest differently across individuals, making accurate diagnosis and treatment highly complex. Even within the same diagnosis, there can be variation in sleep architecture among patients, which makes generalization across people difficult. Traditional sleep analysis methods rely on manual scoring and fixed diagnostic criteria, which fail to capture subject-specific variability in sleep patterns. To address this, we propose a data-driven Personalized Sleep Signature (PSS) approach that learns individualized sleep behaviour using AI models. This study introduces the PSS framework, combining Transformer-based Attention and Graph Attention Networks (GATs) to model nuanced sleep characteristics. We utilize the Nationwide Children’s Hospital Sleep Data, a paediatric Polysomnography (PSG) dataset containing EEG and physiological parameters such as ocular movements, EMG activity, blood pressure, and respiratory rate. From this, we extract sleep epoch features and demographics to form Sleep Signature Groups that reflect common behavioural patterns. Unlike conventional classification, our method captures personal variability and delivers individualized sleep hygiene guidance. The model achieved 94% accuracy in detecting sleep patterns, outperforming traditional methods. Beyond clinical applications, it can be integrated with wearable sensors (e.g., Fitbit, Oura, Apple Watch) to personalize wake/sleep routines and environments. It also enables early detection of sleep disorders and aligns daily schedules with individual chronotypes to enhance well-being. By focusing on sleep behaviour rather than rigid diagnostic categories, this approach supports non-pharmacological, personalized interventions backed by scientific evidence. Our work opens the door to precision sleep medicine, offering actionable insights for clinicians, researchers, and technology innovators.