TabNet based Stress Level Detection using IoT for Real Time Monitoring

Stress is a multifactorial physiological and psychological response that requires reliable real-time monitoring to overcome the limitations of self-report methods and lab-based assessments. In this work, we propose FSIoT-TabNet, an IoT-enabled stress detection framework that integrates multisensor wearable data with fuzzy inference and an enhanced TabNet deep learning model. The system fuses physiological signals (HR/HRV, GSR/EDA, temperature, respiration) and behavioral indicators (activity, sleep patterns) into fuzzy features, which are combined with raw descriptors and processed through TabNet’s sequential attention mechanism. To ensure real-time deployment, the model incorporates modality-specific embeddings, pruning, and quantization for edge optimization. Experiments on the WESAD dataset demonstrate that FSIoT-TabNet outperforms fuzzy-only, machine learning, LSTM, and traditional TabNet baselines, achieving an accuracy of 89.7%, ROC-AUC of 0.92, and the lowest MAE of 4.8, while reducing inference latency to 85 ms on smartphone hardware. These results highlight the model’s ability to deliver both continuous stress indices and categorical stress levels, providing a practical and interpretable solution for real-world stress monitoring.