Determination of candidate predictors for chiropractic treatment outcome of spinal pain using a machine learning framework for small datasets.

OBJECTIVES: To develop a machine learning (ML) approach to explore self-reported factors predictive for recovery in a small set of spinal pain patients. METHODS: In this prospective cohort study, patients (N = 96; mean age = 44.5 ± 16.5 years; 53 female) completed an extensive questionnaire at baseline and after 1 and 3 months. Prediction targets were defined as binary outcomes (recovery/non-recovery) based on improvement at 3 months for pain intensity, disability, quality of life, and Patient Global Impression of Change. The ML-approach included three steps: selection of candidate baseline features (SHapley Additive exPlanations, SHAP); predictor validation (Leave-One-Out Cross-Validation, LOOCV) and permutation testing; testing for the ability to generalize. RESULTS: Area under the curve (AUC) values were 0.93-0.99 in LOOCV and 0.62-0.90 in ensemble cross-validation (CV). SHAP analyses revealed higher recovery odds with positive treatment expectations and higher self-efficacy, younger age, lower body mass index, and fewer comorbidities. Psychological dysfunction generally hindered recovery. CONCLUSIONS: This explorative study suggests that the current ML framework may identify candidate predictors of chiropractic treatment outcome in spinal pain from a small but phenotypically rich dataset. Given the performance drop between LOOCV and ensemble CV, current findings are hypothesis-generating. Prospective replication in adequately powered cohorts is necessary.