Portal de Periódicos da CAPES

Combining Multiple Tiny Machine Learning Models for Multimodal Context-Aware Stress Recognition on Constrained Microcontrollers

2023; Institute of Electrical and Electronics Engineers; Volume: 44; Issue: 3 Linguagem: Inglês

10.1109/mm.2023.3329218

1937-4143

Michael Gibbs, Kieran Woodward, Eiman Kanjo,

Non-Invasive Vital Sign Monitoring

As stress continues to be a major health concern, there is growing interest in developing effective stress management systems that can detect and mitigate stress. Deep Neural Networks (DNNs) have shown their effectiveness in accurately classifying stress, but most existing solutions rely on the cloud or large obtrusive devices for inference. The emergence of tinyML provides an opportunity to bridge this gap and enable ubiquitous intelligent systems. In this paper, we propose a context-aware stress detection approach that uses a microcontroller to continuously infer physical activity to mitigate motion artifacts when inferring stress from heart rate and electrodermal activity. We deploy two DNNs onto a single resource-constrained microcontroller for real-world stress recognition, with the resultant stress and activity recognition models achieving 88% and 98% accuracy respectively. Our proposed context-aware approach improves the accuracy and privacy of stress detection systems while eliminating the need to store or transmit sensitive health data.