Subject-Independent Thermal Infrared Intoxication Detection: Periorbital Deep Learning with Explainability Analysis

Breathalyser based alcohol screening in aviation covers only a fraction of safety-sensitive personnel on any given day, leaving a substantial gap in workforce monitoring. Thermal infrared imaging could help fill this gap, since alcohol consumption raises skin temperature through peripheral vasodilation, and these changes can be picked up passively without any contact. In this work, we apply deep learning to thermal facial images from the University of Patras Sober-Drunk dataset and evaluate performance using Leave-One-Subject-Out Cross-Validation (LOSO-CV), which ensures that the model is always tested on individuals it has never seen during training. Prior studies in this area used subject-dependent evaluation and reported accuracies of 86–98%, but such protocols are known to overestimate performance. Under our subject-independent protocol, the periorbital region proved to be the most informative imaging site (MCC = 0.537, accuracy = 76.2%), and Grad-CAM visualisations confirmed that the model focuses on the periorbital vasculature, which is consistent with where alcohol-induced vasodilation is physiologically expected. We also found that Gaussian noise augmentation, commonly used in image classification, caused the model to lose its focus on thermal gradients altogether, a form of feature collapse verified through Grad-CAM. Fusing periorbital and facial predictions improved both discrimination (MCC = 0.565) and probability calibration (ECE = 0.156). However, around 30% of subjects remained difficult to classify regardless of the model or hyperparameters used, pointing to underlying physiological variability that no amount of architectural tuning could resolve. We propose a practical deployment model where thermal screening acts as a first-pass filter, and anyone flagged is sent for a confirmatory breathalyser test.