Generative Adversarial Networks for Stress Recognition Using Wearables

With the rising rate of stress-induced mental disorders, it is clearly apparent that there is an urgent need for highly precise and efficient stress recognition models. This research addresses the limitations of existing stress recognition models by suggesting an innovative ensemble approach, which tries to not only increase recognition capabilities but also allow a thorough understanding of the basic mechanisms. The main goal of this research was to create and evaluate a novel ensemble model for stress recognition, while using the complementary strengths of both generative and discriminative learning. To accomplish this, a Conditional Generative Adversarial Network (CGAN) model with a Convolutional Neural Network - Long Short Term Memory (CNN-LSTM) network are integrated. Dataset augmentation and better feature representation were handled by CGAN, while temporal dependencies found in physiological signals’ attributes were captured by using the CNN-LSTM. The ensemble model was K-fold cross-validated for proving its robustness and generalization. Also, SHapley Additive exPlanations (SHAP values) is used to enhance model interpretability by informing the user about the importance of the features considered by the model and the reasons behind the decisions made. The results showed a drastic improvement in the accuracy for stress recognition when compared with other state-of-the-art existing methods. The ensemble model had proven its accuracy, precision, and recall to be very high, thereby distinguishing among different classes for stress recognition successfully. The comparative analysis proved that this proposed ensemble model outperforms existing models with better accuracy and robustness; thus, it comes forward to revolutionize this field of stress recognition. Future work will involve modeling the performance along real-life scenarios and implementing it on wearable devices for stress monitoring.