A Hybrid Deep Learning and Ontology-Based Framework for Contextual Hyperactivity Detection in Children with ADHD

International audience We propose a hybrid approach combining deep learning and deductive reasoning through ontology context design to enhance context-aware detection of hyperactivity in children with ADHD. The framework leverages a generative Transformer model through an encoder–decoder architecture enhanced with multi-resolution attention to learn the distribution of normal physical activity patterns. Statistical modeling is then applied to estimate the parameters of the reconstruction error distribution, with deviations beyond a learned threshold classified as anomalies. These detected anomalies are subsequently instantiated in an ontology, which filters them based on context through a set of logical rules. We evaluate our approach on data collected from two children diagnosed with ADHD, recorded under free-living conditions using the Actigraph GT9X device. Our experimental results show average F1-scores of 0.87 and 0.85. These results highlight the potential of our hybrid approach for contextual hyperactivity detection in children with ADHD.