Real-time High-Level Pain Quantification using a Smartphone and a Wrist-worn Electrodermal Activity Sensor (Preprint)

BACKGROUND The subjectiveness of pain leads to inaccurate pain management, which can exacerbate drug addiction and overdose. The consequence is tremendous cost to society and individuals as the opioid crisis grows. Given that pain is often experienced in patients’ homes, there is an urgent need for ambulatory devices that can quantify pain in real time. OBJECTIVE We developed a smartphone-based system for real-time objective pain measurement and assessment using a wrist-worn electrodermal activity (EDA) device. METHODS Our smartphone application collects EDA signals from a wrist-worn device and evaluates pain based on the computation of three pain-sensitive EDA indices: the time-varying index of EDA (TVSymp); modified TVSymp (MTVSymp), and the derivative of phasic EDA (dPhEDA). For testing of our computational algorithms that were embedded in a smartphone application, ten subjects underwent heat pain using a thermal grill, which delivered a level of pain that was calibrated for each subject to be 8 out of 10 on a visual analog scale (VAS). The thermal grill induces heat pain perception without tissue injury using the temperature difference between warm and cold water. All of the wearable-collected EDA signal processing was performed using a smartphone application. Furthermore, another group of fifteen subjects underwent pain stimulation using electrical pulses (EP), which elicited a VAS pain score level 7 out of 10. For EP data collection, EDA signals were collected using a non-wearable device but the same smartphone application was used to calculate the EDA-derived pain indices. We set 5-second segments before and after each pain stimulus to be painless and pain segments, respectively, and trained eight machine-learning classifiers to test the feasibility of our smartphone and EDA-based system to detect pain in real-time. Parameters of the classifiers were optimized using the grid search cross-validation technique. We trained and tested the classifiers on both datasets with leave-one-subject-out cross-validation approach to prevent over-fitting of the models. RESULTS We obtained up to 82.1% accuracy in detecting pain. We also trained using only one dataset at a time and tested with other datasets (and vice versa) and achieved up to 83.1% accuracy. CONCLUSIONS Our results show the potential of a smartphone application to provide near real-time objective pain detection. This approach can potentially enable pain quantification for both acute and chronic pain and it is especially suited for subjects with communication disorders as well as infants.