Predicting psychological resilience and mental health from multimodal wearable sensor data using graph neural networks

Early diagnosis of mental vulnerability is still a problem since most clinical evaluations are based on the subjective aspects of evaluation like self-report and not on the objective physiological indications. The idea of psychological resilience which is a fundamental protective factor against stress related disorders has been connected to the autonomic regulation and daily behavioral patterns, but the objective method of calculating on such a scale is yet to be established. This paper presents a multimodal graph neural network, which combines the wearable derived physiological signals to forecast trait level resilience and stress as well as state level. The ongoing heart activity, electrodermal activity, and motion data of adult participants were gathered and matched with resilient score proven validated scores. Participant specific graphs were created whereby physiological modalities were the nodes and the inter signal dependencies were the edges. Statistical analysis showed that there were significant physiological disparities that were related to resilience. High resilience group members had much better values of root mean square of successive differences and physical activity per day (mean difference on 1000 extra steps per day, p = 0.003). The proposed graph neural network (GNN) performed well in classification tasks in terms of area under the receiver operating characteristic curve (0.80 +- 0.02) to differentiate high and low resilience, which was significantly better than that of logistic regression (0.60), random forest (0.65), and long short-term memory models (0.72), with the difference being supported by DeLong test (p < 0.01). Graph based learning also offers added benefits of discrimination and physiologically explainable output and thus can facilitate its future as a scalable and objective digital mental health monitoring and early risk stratification tool.