Hybrid Transformer-LSTM Architecture for Multimodal Stress Detection Using Physiological Signals

Abstract

Stress significantly affects physical and mental health, motivating the development of automated stress monitoring systems using wearable technologies. Traditional machine learning approaches rely on handcrafted features that often fail to capture complex temporal dependencies in physiological signals. This paper proposes a hybrid Transformer-LSTM framework for multimodal stress recognition using Heart Rate Variability (HRV), Electrodermal Activity (EDA), skin temperature, respiration rate, and accelerometer data. The Transformer encoder captures long-range temporal relationships, while the LSTM network models sequential physiological variations. An attention-based fusion mechanism integrates heterogeneous sensor modalities to improve classification performance. Extensive experiments demonstrate superior accuracy compared to conventional deep learning architectures, including CNN, LSTM, GRU, and Bi-LSTM models. The proposed framework provides a robust solution for continuous stress monitoring in healthcare, workplace wellness, and mental health assessment applications.