AI-Based Risk Stratification and Explainable Models for Wearable Mental Health Monitoring Systems

Wearable artificial-intelligence systems using explainable-AI methods have been proved effective for neurological disorders like epilepsy, Parkinson's disease, stroke and Alzheimer's disease. Adapting the same concepts to depression, anxiety, and stress is not as easy; meta-analytic findings show that mood disorders identification largely continues to base itself on one public dataset only, prefer past studies with very few studies extending outside of the wrist area, and hardly any tests beyond the wrist have been done. We highlight these issues and propose a multimodal sensing and modeling scheme that incorporates photoplethysmography, electrodermal activity, actigraphy, skin temperature, and sleep stages with the help of a CNN-BiLSTM-Attention model which is the extension of CNN-BiLSTM-Attention and XAI system already tested on CHB-MIT scalp EEG for predicting seizures to biosignals concerned with mental health. Continuous physiological measurements open an opportunity for hacking as well. So, besides the modeling component, we include a secure-transmission five-layer stack adapted from a wireless-sensor-network one tested within a military body-area-network environment. We wrap up this piece with a clinical effectiveness, explainability consistency, and transmission security plan under a simulated breach which aims at closing the specific gaps addressed by our design.