Contactless Sleep Monitoring with Potential Application to Early Autism Detection

patients, diagnosing chronic conditions, and enable behavioural detection of autism. Existing methods, particularly visionbased and wearable systems, face limitations including privacy concerns, sensitivity to lighting and occlusion, complexity in handling long video sequences, and neurodivergent users discomfort. To address these challenges, this study proposes a privacy-preserving, contactless sleep monitoring system that leverages ultra-wideband (UWB) radar and lightweight deep learning (DL) architectures. The system performs multi-class micro-Doppler-based classification of nine fine-grained sleep activities: body left, body right, feet move, hand move, head left, head right, static head left, static head right, and static head up. These postures are of relevance in identifying subtle behavioural indicators associated with early-stage autism, offering potential for preclinical behavioural screening. A comparative evaluation is conducted using DL models, VGG16, VGG19, MobileNet, and SqueezeNet, on radar-derived datasets. Among these, the VGG16 model achieves the highest classification accuracy of 84.6% on the combined dataset. The results confirm the feasibility of deploying accurate and efficient models in real-world, low-resource healthcare settings, without compromising user privacy or comfort.