Individualized machine-learning predictions of pain intensity in chronic low back pain

Chronic low back pain (cLBP) imposes a major public health burden and its presentation is further complicated by high comorbidity with depression. Processes that maintain cLBP are likely highly individualized and span biological, psychological, and social domains, posing significant challenges for optimal treatment. This proof-of-concept study applied a systematic individualized machine learning (ML) pipeline to predict individual pain intensity from intervenable biopsychosocial variables obtained from ecological momentary assessment and smartwatch wearables. Participants were six adults with cLBP and major depressive disorder (Mage=43.0, SD=15.14; 67% female). In total, 29 candidate predictor variables were measured daily for 30 days, including metrics of physical activity, sleep, heartrate, negative and positive mood, diet quality, and social connection. We fit ML models per participant and determined each participant's top-ranked variables predicting pain intensity. Results revealed models with an average error of 28.4% +/- 12.3% across participants. The top predictors of pain intensity were largely unique to each participant. For example, whereas depressed and anxious mood were within the top five predictors of pain intensity for the majority of participants, predictive strengths varied considerably across participants. We also found notable differences in predictive strengths for positive mood, physical activity, diet quality, and social factors. This study provides proof-of-concept of our systematic individualized modeling ML pipeline for identifying individual, intervenable predictors of pain intensity in adults with cLBP and depression. Future research is required for additional development and validation of the ML pipeline and for its utility for guiding personalized treatment strategies. PERSPECTIVE: Chronic low back pain is highly prevalent and varies across individuals. Our findings demonstrate proof-of-concept of applying a systematic individualized machine learning pipeline to predict individual pain intensity from intervenable biopsychosocial variables. This approach may provide a foundation for future research examining more personalized approaches to pain assessment and management.