A classification algorithm to predict chronic pain using both regression and machine learning - A stepwise approach.

The sample was small, and the number of participants available for each weekly analysis varied from 43 to 59. The study used only step count and sleep hours as sensor parameters, so other pain cues were not tested. Few participants had moderate or higher pain, so the models were built on a limited number of pain cases. Pain was recalled weekly and different pain types were lumped together, which may not capture daily pain changes. Sleep hours came from a fitness tracker, not polysomnography, and may be less accurate. Mood was measured with two questions, not a full depression scale. Participants were mostly white, low income, and had no dementia, so the findings may not apply to people with severe cognitive or communication impairment. The authors reported that the pain intensity model needed better positive predictive value and the pain interference model had poor sensitivity. The paper describes future use, not a clinical test of the model.

The supplied metadata lists NIH extramural and U.S. government non-PHS research support. No author conflict-of-interest statement is provided.

Do not conclude that a wearable step-count app is ready to screen or diagnose pain, or that increasing steps will reduce pain. This was a pilot prediction analysis in 77 older adults, not a treatment study. The authors also said the pain intensity model needed improvement, and the pain interference model had poor sensitivity, meaning it may miss patients with pain interference.