Prediction of pain outcomes in a randomized controlled trial of dose–response of spinal manipulation for the care of chronic low back pain

Darcy Vavrek, Mitchell Haas, Moni Blazej Neradilek, Nayak Polissar

No previous studies have created and validated prediction models for outcomes in patients receiving spinal manipulation for care of chronic low back pain (cLBP). We therefore conducted a secondary analysis alongside a dose-response, randomized controlled trial of spinal manipulation. We investigated dose, pain and disability, sociodemographics, general health, psychosocial measures, and objective exam findings as potential predictors of pain outcomes utilizing 400 participants from a randomized controlled trial. Participants received 18 sessions of treatment over 6-weeks and were followed for a year. Spinal manipulation was performed by a chiropractor at 0, 6, 12, or 18 visits (dose), with a light-massage control at all remaining visits. Pain intensity was evaluated with the modified von Korff pain scale (0-100). Predictor variables evaluated came from several domains: condition-specific pain and disability, sociodemographics, general health status, psychosocial, and objective physical measures. Three-quarters of cases (training-set) were used to develop 4 longitudinal models with forward selection to predict individual "responders" (≥50 % improvement from baseline) and future pain intensity using either pretreatment characteristics or post-treatment variables collected shortly after completion of care. The internal validity of the predictor models were then evaluated on the remaining 25 % of cases (test-set) using area under the receiver operating curve (AUC), R(2), and root mean squared error (RMSE). The pretreatment responder model performed no better than chance in identifying participants who became responders (AUC = 0.479). Similarly, the pretreatment pain intensity model predicted future pain intensity poorly with low proportion of variance explained (R(2) = .065). The post-treatment predictor models performed better with AUC = 0.665 for the responder model and R(2) = 0.261 for the future pain model. Post-treatment pain alone actually predicted future pain better than the full post-treatment predictor model (R(2) = 0.350). The prediction errors (RMSE) were large (19.4 and 17.5 for the pre- and post-treatment predictor models, respectively). Internal validation of prediction models showed that participant characteristics preceding the start of care were poor predictors of at least 50 % improvement and the individual's future pain intensity. Pain collected shortly after completion of 6 weeks of study intervention predicted future pain the best.