NANS 2021

Predicting Pain Now and in the Future Through Personalized Physiologic Mobility Metrics


Introduction: Objectively detecting and grading pain, particularly chronic pain, has remained an elusive goal This is in part due to its subjective and multivariate nature, and the inherent limitations in measuring and under- standing pain experience at the individual level Currently, patient self reported scales of pain intensity (VAS/NRS/ serve as gold standard measures in pain management, despite known inaccuracies and bias of reporting in these measures Novel objective metrics are needed to better capture patient experience, avoid bias, and improve repeatability in order to inform physi- cians and patients of treatment options These metrics need to describe physiology, capture the range of changes during treatment, and ideally be suited to predicting future changes so that interventions can be taken. Methods: Through on-going, multi-site Boston Scientific studies of spinal cord stimulation (SCS) response in chronic leg and back pain patients, we monitored self-reported pain and patient activity. These studies will involve up to 1700 patients at up to 30 sites. Patients used smartphone apps to answer questionnaires (2x/day) and wore smartwatches daily. Fifty-three patients with overlapping data were analyzed. We developed novel artificial intelligence methods to derive personalized patient models that charac- terize current pain and predict pain over the coming days. This approach clusters patients into homogeneous groups based on their pain profiles and leverages notions of cluster similarity to model pain. The modeling is done in three stages, starting with a learning representation implemented with neural network layers to produce embeddings for patient data. Next, the embeddings are used with historical pain data by a Dynamic-Time-Warping inspired clustering algorithm to group patients in homogeneous pain clusters. We then learn personalized pain models in a multi-task-learning setting, using additional neural network layers. Key to this analysis was the engineering of a custom temporal feature from smartwatches (Effective- Mobility) to capture patient activity. Results: Initial results demonstrate the effectiveness of the proposed pain estimation and prediction method (Table-1/Figure-2). Results show personalized models based on objective smartwatch data perform similarly to those of models based on subjective questionnaires alone. Also, personalized models outperform population-based models. Conclusion: Our preliminary results demonstrate an individual patient’s pain can be captured from objective smartwatch data alone using novel AI techniques. Further, the techniques can also predict the patient’s pain for the next day. We found personalized models to be more predictive than population ones for such tasks and can provide clinically relevant performance.