Interpretable Machine Learning Models for Short- and Long-term Prognostic Prediction and Risk Factor Identification in Radiofrequency Treatment of Lumbar Facetogenic Pain: A Retrospective Cohort Study with Temporal Validation.

Abstract

Study design: Retrospective cohort study.

Objective: To develop machine learning (ML) models integrating clinical/imaging variables for predicting 3- and 6-month outcomes of radiofrequency (RF) treatment in lumbar facetogenic pain, and an independent temporal validation cohort was used to evaluate the model's performance. Shapley Additive Explanations (SHAP) analysis was utilized to identify key variables and construct a simplified model.

Summary of background data: Early identification of RF-responsive patients remains challenging, with limited non-invasive prognostic tools available.

Methods: Six ML models were trained using 16 clinical/imaging variables from 372 RF-treated patients. Model performance was evaluated via AUROC, with SHAP analysis identifying key variables. Simplified models using clinical-only, imaging-only, and SHAP-selected variables were compared.

Results: In the discovery (n=312) and temporal validation (n=60) cohorts, 141 and 26 patients had unsuccessful 3-month outcomes, respectively. The logistic model outperformed others, achieving AUROCs of 0.834 (95% CI: 0.725-0.942) and 0.818 (0.713-0.923) for 3-month prediction in discovery and validation cohorts. Simplified models showed comparable performance (discovery AUROC: 0.795-0.837; validation: 0.699-0.814). Six-month predictions demonstrated similar robustness (discovery AUROC: 0.813; validation: 0.783). Decision curve analysis confirmed the logistic model's clinical utility, providing net benefits at threshold probabilities >40%.

Conclusions: The Logistic model, which is based on clinical and imaging variables, has the potential to facilitate early screening of patients who might benefit from RF treatment in the short- and long-term. SHAP analysis helps evaluate the impact of variables and build simplified models with comparable performance. The key variables identified in this study require further verification through external geographic validations.

Level of evidence: 3.