PainRhythms: Machine learning prediction of chronic pain from circadian dysregulation using actigraph data — a preliminary study

Chronic pain is currently diagnosed using verbal self-reports, which present a challenge for patients with cognitive or physiological disorders. Prior work has explored machine learning prediction of pain from clinical data, which requires active user involvement and does not capture their behavior in natural settings. Passive objective assessment is desirable. Circadian Rhythms, including sleep–wake cycles, are biological processes that reoccur every 24 h and can be derived from physiological data such as heart rate, activity, and sleep, gathered using widely-owned smart wearables. This study investigated the feasibility of using machine learning and rest-activity circadian rhythm features to predict patients’ pain, including pain intensity, its interference with the patient’s life (dysregulation), and their difficulty in performing physical functions using passively gathered actigraphy data. To predict pain on day N, actigraphy data collected over that day were analyzed. Three sets of feature were extracted: (1) Activity (total sedentary bouts/time/breaks, % in sedentary/light/ moderate activity), (2) Sleep (sleep efficiency/latency, wake after sleep onset), and (3) Rest Activity Rhythm (mesor, acrophase, Intradaily Variability (IV)). These features were then classified using various machine learning algorithms. Our proposed PainRhythms approach achieved an average AUC-ROC of 0.97 with a stacking machine learning classifier for predicting pain, 0.67 and 0.62 with logistic regression for pain intensity and interference, and 0.56 with gradient boosting for physical function. We found that chronic pain predictions were more accurate using rest-activity rhythm features than sleep or activity features. Of all the rhythmic features, Intradaily Variability (IV) was the most predictive feature, with elevated values in pain associated with disturbed sleep. PainRhythms provides preliminary evidence that rest-activity rhythms can effectively detect subjects with chronic pain. In future work, we aim to gather more data and confirm our preliminary findings on a large, class-balanced and diverse dataset.