A Fusion Model for Cross-Subject Stress Level Detection Based on Transfer Learning

Abstract

Stress is a psychological condition that affects daily life, and chronic stress can result in cardiovascular disease and reduced productivity. Mental stress can be induced when difficult and time-limited tasks are assigned. Several groups have studied the relationship between physiologic signals and a subject's stress level. Through machine learning and signal processing, stress level can be automatically inferred from raw physiologic signals. As each person can have a specific physiologic reaction pattern to stress, it becomes problematic for a classifier to work well on a new subject. In this study, transfer learning is used to solve the problem of inter-subject variability. Methods are developed here to classify five levels of stress based on physiologic signals comprising photoplethysmogram (PPG), galvanic skin response (GSR), abdominal respiration, and thoracic respiration. Domain adaptation methods based on information-theoretical learning and transfer component analysis (TCA) are shown to reduce inter-subject variability of both GSR and respiratory signals. A fusion model was also designed to combine classification scores from each signal to reduce the effect of low-quality recording. The proposed method is shown to increase accuracy from 68.79% to 76.70% and Intraclass Correlation Coefficient (ICC) from 83.82% to 96.55%.