Toward Detecting and Explaining Stress of Nurses Using Wearable Devices and LLMs

In this paper, we improve nurse stress detection using wearable device data by balancing the dataset through Large Language Models (LLMs). The rise in wearable technologies has spurred stress detection studies in both daily life and healthcare settings. However, integrating an LLM on stress detection with wearable devices is still emerging. Data imbalance and faulty sensor readings, especially from EDA sensors in devices like the E4, can significantly impact the accuracy of stress detection models. To address this, we utilize LLMs to mitigate data imbalance by generating relevant synthetic samples through varied contextual prompting strategies. A Random Forest (RF) classifier is trained using physiological sensor datasets including acceleration, electrodermal activity (EDA), heart rate (HR), and skin temperature (Temp). We compared Zero-shot, Few-shot, Zero-shot CoT, and Few-shot CoT prompting ability on data generation. We benchmark the LLM approach against SMOTE and Random Sampling comparing overall performance on stress detection. Furthermore, we employ the RACCCA (Relevance, Accuracy, Completeness, Clarity, Coherence, Appropriateness) framework to assess the quality of prompting and output. Overall, the proposed data augmentation using LLMs improved performance compared to the baseline. Specifically, Zero-Shot prompting increased the overall F1-score by 10% (from 88.12% to 98.11%). This study underscores the potential of LLM-generated synthetic data in addressing data imbalance and enhancing stress detection models.