Background: Recent studies on machine learning have shown the potential to provide new methods with which to assess pain through the measurement of signals associated with physiologic responses to pain detected by wearables. We conducted a prospective pilot study to evaluate the real-world feasibility of using an AI-enabled wearable system for pain assessment with elderly patients with dementia and impaired communication.

Methods: Sensor data were collected from the wearables, as well as observational data-based conventional everyday interventions. We measured the adherence, completeness, and quality of the collected data. Thereafter, we evaluated the most appropriate classification model for assessing the detectability and predictability of pain.

Results: A total of 18 patients completed the trial period, and 10 of them had complete sensor and observational datasets. We extracted 206 matched records containing a 180 min long data segment from the sensor's dataset. The final dataset comprised 153 subsets labelled as moderate pain and 53 labelled as severe pain. After noise reduction, we compared the recall and precision performances of 14 common classification algorithms. The light gradient-boosting machine (LGBM) classifier presented optimal values for both performances.

Conclusions: Our findings tended to show that electrodermal activity (EDA), skin temperature, and mobility data are the most appropriate for pain detection.