Real-time Evaluation of an Automated Computer Vision System to Monitor Pain Behaviour in Older Adults

Abstract

A large body of literature supports the systematic observation of facial expressions as a tool for assessing pain in both younger and older adults. Such observation is especially critical for older adults who have limited ability to communicate their pain experience due to dementia. While frequent monitoring of pain behaviours in dementia is constrained by resource limitations, computer vision technology has the potential to mitigate these challenges, especially in long-term care environments where many people with severe dementia reside. A computerized algorithm designed to assess pain behaviour in older adults with and without dementia was recently developed and validated against video recorded images. The algorithm was incorporated within an automated system that provided alerts when facial pain expressions were detected. This study conducted the first live, real-time evaluation of the automated pain behaviour detection system with community-dwelling older adults in a laboratory setting. Testing involved a total of 65 participants completing three safely-administered experimentally-induced pain tasks using a thermal pain stimulator. A video camera was used to facilitate recording and automatic processing of facial activity. Pain behaviour detection occurred when systemgenerated pain intensity scores of the facial expressions displayed by participants exceeded a predetermined threshold score. When the incidence of facial pain expression occurred, an electronic notification (e.g., email and a signal light) was generated as notifications of pain behaviour detection. Participants completed continuous self-report pain intensity ratings during the thermal pain tasks. Receiver Operating Characteristic curve analyses were used to determine the sensitivity and specificity of the system in identifying pain- and non-pain facial expressions using gold standard manual coding completed by trained coders. Gender differences were also explored in relation to system performance. Correlational procedures were used to evaluate the relationship between pain intensity scores generated by the system, continuous self-report pain ratings, observational pain coding, and stimulus temperatures. This study supported the potential viability of the automated pain behaviour detection system in correctly identifying live, real-time instances of facial pain expressions in older adults. System-generated pain behaviour scoring achieved a maximal greater correlation with gold standard manual coding compared to prior testing using video-recordings. Specifically, system performance improved when more intense facial pain expressiveness was considered compared to more subtle facial expressions at lower pain intensities. In comparing system scoring to manual coding, there was not a one-to-one correspondence in coding but a range of comparative values that varied from participant to participant. Correlational analyses showed that continuous self-report pain ratings were weakly correlated with system scoring and manual coding. While average pain scores remained homogenous across genders, results suggested that the system performed better at identifying pain expressions for women compared to men. As expected, the pain-related facial movements of brow lowering and levator contraction were unique predictors of system-generated scores. Future evaluations of the system involving field trials in long-term care settings with older clinical populations would further elucidate the performance of the system. This technology is expected to aid in the assessment of pain in people living with dementia while addressing resource constraints in long-term care environments and reduce burden for caregivers. Keywords: Pain, aging, technology, older adults, computer vision, dementia