Quantum-entangled feature selection and spiking graph transformer networks for early detection of childhood behavioral markers.

Introduction: Early identification of childhood behavioural markers using wearable sensing is important for timely intervention in developmental and sleep-related disorders. Wrist-worn accelerometer data provide objective measures of behavioural regulation by capturing actigraphy-derived states such as Sleep, Wake, and Transitional periods. However, existing deep learning methods for behavioural state detection often face challenges related to redundant features, sensitivity to sensor noise, limited robustness in long-term wearable deployment, and high energy consumption. Methods: This study proposes a Quantum Variational Feature Selection-Spiking Graph Transformer Network (QVFS-SGTN), a hybrid quantum-neuromorphic framework for robust and energy-efficient behavioural state classification. The prediction task was defined as a three-class classification problem involving Sleep, Wake, and Transitional behavioural states from high-frequency wrist-worn accelerometer data. The proposed model integrates a parameterized quantum circuit-based feature selector to identify non-linear and entangled sensor correlations. Selected features are then processed by a spiking graph transformer network, which models temporal dependencies through event-driven self-attention and neuromorphic neuron dynamics. Experiments were conducted using the Child Mind Institute wearable dataset, with additional cross-dataset validation performed on the external Multi-Ethnic Study of Atherosclerosis sleep dataset. Results: The proposed QVFS-SGTN framework achieved state-of-the-art performance on the Child Mind Institute wearable dataset, with an accuracy of 0.968, an F1-score of 0.968, and an AUC of 0.991. Robustness evaluation demonstrated stable performance under significant Gaussian noise, maintaining accuracy above 92%. Energy analysis showed a 40-55% reduction in computational cost compared with conventional deep learning models. In cross-dataset evaluation using the MESA sleep dataset, the model achieved an accuracy of 0.931 without fine-tuning, indicating strong generalization capability. Discussion: The findings demonstrate that combining quantum-enhanced feature selection with spiking graph-based temporal modelling can improve the robustness, accuracy, and energy efficiency of wearable behavioural state detection. The QVFS-SGTN framework effectively addresses key limitations of existing deep learning approaches, including feature redundancy, sensor noise sensitivity, and computational cost. These results support the potential of the proposed hybrid quantum-neuromorphic model for scalable, long-term, real-world paediatric behavioural monitoring.