Deployment of TinyML-Based Stress Classification Using Computational Constrained Health Wearable

Stress has become a common mental health issue in modern society, causing individuals to experience acute behavioral changes. Exposure to prolonged stress without proper prevention and treatment may cause severe damage to one’s physiological and psychological health. Researchers around the world have been working to find and create solutions for early stress detection using machine learning (ML). This paper investigates the possibility of utilizing Tiny Machine Learning (TinyML) in developing a wearable device, comparable to a smartwatch, that is equipped with both physiological and psychological data detection system to enable edge computing and give immediate feedback for stress prediction. The main challenge of this study was to fit a trained ML model into the microcontroller’s limited memory without compromising the model’s accuracy. A TinyML-based framework using a Raspberry Pi Pico RP2040 on a customized board equipped with several health sensors was proposed to predict stress levels by utilizing accelerations, body temperature, heart rate, and electrodermal activity from a public health dataset. Moreover, a few selected machine learning models underwent hyperparameter tuning before a porting library was used to translate them from Python to C/C++ for deployment. This approach led to an optimized XGBoost model with 86.0% accuracy and only 1.12 MB in size, hence perfectly fitting into the 2 MB constraint of RP2040. The prediction of stress on the edge device was then tested and validated using a separate sub-dataset. This trained model on TinyML can also be used to obtain an immediate reading from the calibrated health sensors for real-time stress predictions.