Predicting short-term disability outcomes in low back pain using baseline patient-reported measures and machine learning models

Predicting short-term outcomes for low back pain (LBP) based on baseline characteristics remains challenging. Machine learning (ML) has the potential to improve outcome prediction by integrating multidimensional clinical data. To examine whether baseline patient-reported outcome measures (PROMs) and functional assessments predict short-term rehabilitation outcomes in nonspecific LBP using supervised ML. We analyzed 100 adults with nonspecific LBP completing a four-week program. Five algorithms (logistic regression, decision tree, random forest, support vector machine, XGBoost) were trained using baseline PROMs (Oswestry disability index [ODI], Roland-Morris Disability Questionnaire [RMDQ], Quebec disability index [QDI], short form-12 [SF-12]) and functional measures (hip mobility, muscle force, functional movement). Performance was evaluated via five-fold cross-validation for classifying treatment responders (≥30% ODI reduction) and predicting post-intervention ODI scores. For responder classification, random forest performed best (AUC=0.68, accuracy=71%, sensitivity=85%). In regression analysis, it explained the most variance (R2=0.28, RMSE=7.2). Feature importance revealed PROMs (SF-12, QDI, RMDQ) were the most influential predictors, whereas objective functional measures contributed less. Decision curve analysis showed a positive net benefit across threshold probabilities of 0.2-0.5. These findings demonstrate preliminary feasibility for applying machine-learning approaches to routinely collected PROMs in short-term rehabilitation outcome prediction in low back pain. However, predictive performance remained modest, and the models should be considered exploratory and hypothesis-generating pending external validation in larger multicenter cohorts.