Physiological Signal Data-Driven Workplace Stress Detection Among Healthcare Professionals Using BiLSTM-AM and Ensemble Stacking Models: Fid-Bau Portal

Physiological Signal Data-Driven Workplace Stress Detection Among Healthcare Professionals Using BiLSTM-AM and Ensemble Stacking Models

Pasha, Syed Tangim / Halder, Nabarun / Badrul, Tasnuba / Setu, Jahanggir Hossain / Islam, Ashraful / Alam, Md Zahangir

Abstract

The multifaceted nature of workplace stress in the healthcare sector is compounded by the physiological variances inherent to each individual. The challenge is further augmented by the lack of precisely labeled data for accurate stress detection through biosignals. Timely identification of stress in these settings is crucial, as it enables healthcare professionals to adopt effective management strategies, thereby improving their overall well-being and maintaining the quality of patient care. Advanced wearable technology e.g., smartwatches has boosted momentum to diagnose stress because these gadgets can gather physiological data in real-time. Our study utilized a publicly available multimodal nurse dataset collected through the Empatica E4 smartwatch. The dataset consisted of multi-class (no stress (‘0’), low stress (‘1’), and high stress (‘2’)) labels. We used a prepro-cessed dataset that was formed of three specific physiological signals, i.e., Heart Rate (HR), Electrodermal Activity (EDA), and Skin Temperature (ST). Moreover, the preprocessed nurse dataset is composed of 48 features. For nurse stress detection, we developed a Bidirectional Long Short-Term Memory with Attention Mechanism (BiLSTM-AM) model, and a stacking-based Ensemble model forming of Decision Tree (DT), Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Multi-Layer Perceptron (MLP), and Logistic Regression (LR). Our BiLSTM-AM and the Ensemble model surpassed most of the state-of-the-art model performance, and they achieved accuracies of 96 % and 97 % respectively. Furthermore, we evaluated the model's performance using precision, recall, F1-score, accuracy, confusion matrix, and Receiver Operating Characteristic (ROC) curve metrics. We also performed an ablation study on our proposed BiLSTM-AM model to enhance the robustness of the architecture for the stress classification task. We believe this study will contribute to developing an automated stress detection system using wearable sensors that will improve the mental well-being of other healthcare professionals, as well as nurses.