Noninvasive and Objective Near Real-Time Detection of Pain Changes in Fluctuating Pain

Abstract

Chronic pain involves natural intensity fluctuations that patients cannot control, contributing to learned helplessness and functional impairment. Detecting spontaneous pain decreases could enable precisely timed interventions that enhance perceived control. However, existing research on objective pain assessment has focused primarily on estimating static intensity from short, predictable stimuli rather than detecting moment-to-moment changes in ongoing pain. This study investigated whether pain decreases can be detected objectively using easily obtainable physiological signals during fluctuating pain. We recorded multiple physiological signals (8-channel EEG, electrodermal activity, heart rate, pupil diameter, facial expressions) from 42 healthy participants (Mage = 26.2 years, SD = 5.1) during calibrated tonic noxious heat stimulation on the left forearm with unpredictable intensity fluctuations (0-70 on visual analogue scale; twelve 3-minute trials). Temperature changes lasted 5-20 seconds. Using minimal preprocessing suitable for real-time applications, we trained deep learning models to classify pain decreases versus non-decreases from brief temporal windows, evaluated on a held-out test set (9 participants). Combining electrodermal activity, heart rate, and pupil diameter yielded optimal classification performance using a transformer-based architecture (AUROC = 0.854, accuracy = 76.8%). Electrodermal activity emerged as the most informative single predictor. Continuous stream analysis demonstrated median detection latency of 5.75 seconds with 70.4% sensitivity, reducible to 4.25 seconds at the cost of increased false positives. Results indicate that electrodermal activity and heart rate enable straightforward practical deployment, while highly variable signals such as EEG and facial expressions require personalized fine-tuned models. These findings establish a basis for closed-loop interventions targeting spontaneous pain changes. SummaryDeep learning models detect pain decreases using electrodermal activity, heart rate, and pupil diameter (AUROC=0.854, 5.75s latency), facilitating precisely timed interventions for fluctuating tonic pain.