BACKGROUND

Neck pain (NP) ranks among the leading causes of years lived with disability worldwide. While spinal manipulation is a common physical therapy intervention for NP, its variable patient responses and inherent risks necessitate careful patient selection. This study aims to develop and validate a machine learning-based prediction model to identify NP patients most likely to benefit from spinal manipulation.

METHODS

This multicenter study analyzed 623 NP patients in a retrospective cohort and 319 patients from a separate hospital for external validation, with data collected between May 2020 and November 2024. Treatment success was defined as achieving ≥ 50% reduction in Numerical Rating Scale (NRS) and ≥ 30% reduction in Neck Disability Index (NDI) after two weeks of spinal manipulation. We compared data imputation methods through density plots, and conducted δ-adjusted sensitivity analysis. Then employed both Boruta algorithm and LASSO regression to select relevant predictors from 40 initial features, and four feature subsets (Boruta-selected, LASSO-selected, intersection, and union) were evaluated to determine the optimal combination. Nine machine learning algorithms were tested using internal validation (70% training, 30% testing) and external validation. Performance metrics included Area Under the Receiver Operating Characteristic Curve (AUC), accuracy, F1-score, sensitivity, specificity, and predictive values. The SHAP framework enhanced model interpretability. Youden's Index was applied to determine the optimal predictive probability threshold for clinical decision support, and a web-based application was developed for clinical implementation.

RESULTS

The combined LASSO and Boruta algorithms identified nine optimal predictors, with the union feature set achieving superior performance. Among the algorithms tested, the Multilayer Perceptron (MLP) model demonstrated optimal performance with an AUC of 0.823 (95% CI 0.750, 0.874) in the test set, showing consistency between training (AUC = 0.829) and test performance. External validation confirmed robust performance (AUC: 0.824, accuracy: 0.765, F1 score: 0.76) with satisfactory calibration (Brier score = 0.170). SHAP analysis highlighted the significant predictive value of clinical measurements and patient characteristics. Based on Youden's Index, the optimal predictive probability threshold was 0.603, yielding a sensitivity of 0.762 and specificity of 0.802. The model was implemented as a web-based application providing real-time probability calculations and interactive SHAP force plots.

CONCLUSION

Our machine learning model demonstrates robust performance in identifying suitable candidates for spinal manipulation among neck pain patients, offering clinicians an evidence-based practical tool to optimize patient selection and potentially improve treatment outcomes.