<![CDATA[Stress is a psychophysiological condition that requires continuous and objective monitoring. However, existing wearable stress detection systems often rely on cloud-based processing or computationally intensive algorithms, limiting their applicability for real-time inference on resource-constrained embedded devices. This study presents a TinyML-based framework for real-time stress detection using the MAX30102 sensor and an ESP32 microcontroller. The proposed framework integrates four time-domain Heart Rate Variability (HRV) features (BPM, SDNN, RMSSD, and pNN50), a lightweight Deep Neural Network (DNN), full INT8 TensorFlow Lite quantization, and on-device inference to enable efficient edge-based stress classification. The DNN model was trained and evaluated using the WESAD dataset. Experimental results showed that a decision threshold of 0.70 yielded the best classification performance, achieving an accuracy of 82% and an F1-score of 0.63 for the stress class. The quantized TensorFlow Lite INT8 model preserved 100% prediction compatibility between the Python and ESP32 implementations. Furthermore, the MAX30102 sensor achieved a BPM measurement accuracy of 98.74%, while the HRV feature extraction implemented on the ESP32 produced results consistent with the reference calculations. These findings demonstrate that the proposed end-to-end TinyML framework enables accurate and computationally efficient HRV-based stress detection on resource-constrained microcontrollers, providing a practical foundation for real-time wearable edge-health monitoring]]>
