A time series machine learning approach for vigilance classification by non-invasive infrared thermography

Abstract

Improvement of vigilance measurement and data analysis is needed to make vigilance classification more applicable in real-world situations. This study aims to evaluate whether non-invasive continuous infrared thermography can be used as a vigilance measurement, whether machine learning models can improve data analysis and whether adding Time Series Analysis can improve vigilance classification. A 10-minute psychomotor vigilance task was conducted with 29 participants. The baseline Generalized Linear Model was compared to a Support Vector Machine model with Radial Basis Function and a Long-Term Short Memory neural network model. Three distal-to-proximal temperature gradients measured by iButtons and an infrared camera were used as predictors to classify vigilance. The Hanley and McNeil test showed no difference between the models and different model predictors. All models classified vigilance around chance level. Future research should include a more distributed participant group and more statistical features.