Finding predictive factors of exercise adherence in randomized controlled trials on low back pain: an individual data re-analysis using machine learning techniques.

Abstract

Objective: To identify predictors of adherence in supervised and self-administered exercise interventions for individuals with low back pain.

Design: Cohort study.

Setting: Rehabilitation.

Participants: This pre-planned re-analysis within the MiSpEx Network included 1,511 participants with low back pain (57% female, mean age 40.9 years, SD ±14 years).

Interventions: Participants underwent an initial 3-week supervised phase of sensorimotor exercises, followed by a 9-week self-administered phase.

Main outcome measures: Biological, psychological, and social factors potentially impacting training adherence were evaluated. During the supervised phase, adherence was tracked through a standardized training log. During the self-administered phase, adherence was monitored via a diary, with adherence calculated as the percentage of scheduled versus completed sessions. Adherence was analyzed both as an absolute percentage and as a dichotomized variable (adherent vs. non-adherent, with a 70% adherence cut-off). Predictors for adherence were identified using Gradient Boosting Machines and Random Forests (R-package caret). Seventy percent of the observations were used for training, while 30% were retained as a hold-out test set.

Results: The average overall adherence was 64% (±31%), with 81% (±28%) adherence during the supervised phase and 58% (±39%) in the self-administered phase. The root mean square error for the test-set ranged from 36.2 (R²=0.18, self-administered phase) to 19.3 (R²=0.47, supervised phase); prediction accuracy for dichotomized models was between 64% and 83%. Predictors of low to intermediate adherence included poorer baseline postural control, decline in exercise levels, and fluctuations in pain intensity (both increases and decreases).

Conclusion: Identified predictors could aid in recognizing individuals at higher risk for non-adherence in low back pain exercise therapy settings.