Wearable Long-Term Graph Learning for Non-invasive Mental Health Evaluation

Long-term non-invasive mental health monitoring based on wearable devices has received increasing attention. In the study, a wearable sensing device that can collect wearer’s audio, behavior, and environmental sensing data in real-time without infringing on the wearer’s privacy is designed. Using this system, we have conducted a 4-week physical and mental health experiment for college students. For audio features, we use full connection and nearest neighbor connection to construct individual graphs, and then develop graph convolutional networks to identify autism. We have also developed an explanatory model to extract subgraphs of individual graphs. The results show that the afternoon or evening time period has a greater impact on high-scoring autistic people.