TinyStressNAS: Automated Feature Selection and Model Generation for On-device Stress Detection

Over the past decades, Electrodermal Activity (EDA)-based stress detection systems have observed increasing interest due to the significant impact of stress on daily life and the need for continuous, unobtrusive monitoring. The wearable form-factor and the unobtrusive nature of the EDA sensor make it highly suitable for this purpose. Many researchers have focused on developing EDA-based stress detection techniques, but few consider the resource footprint of such systems. Maintaining a good trade-off by manual optimization is a challenging task. Automation techniques can be very helpful in such scenarios. In this work, we introduce a novel method for the selection of an optimal feature subset, utilizing a deep Reinforcement Learning (RL) network on top of a Neural Architecture Search (NAS) framework to design tiny, customized models for on-device stress detection. The feature subset selection uses a multi-metric reward objective function to balance feature accuracy contributions against computational complexity, identifying the optimal feature subset. Initial analysis indicates that this optimal subset enhances accuracy while keeping the model size under 100 kB and minimally increasing computation.