Prediction of Therapeutic Outcome after Transforaminal Epidural Steroid Injection in Patients with Radiculopathy following Lumbar Spinal Stenosis Using Machine Learning.

BACKGROUND

Transforaminal epidural steroid injections are widely used to manage radicular pain caused by lumbar spinal stenosis; however, therapeutic outcomes vary among patients. Accurately predicting a transforaminal epidural steroid injection's response may help clinicians optimize treatment planning.

OBJECTIVES

Our study aimed to develop and validate machine learning models, particularly a deep learning neural network, to predict therapeutic outcomes post transforaminal epidural steroid injections using both clinical and magnetic resonance imaging data.

STUDY

DESIGN

A study using machine learning analysis.

SETTING

The spine center of a university hospital.

METHODS

A retrospective dataset of 188 patients with lumbar spinal stenosis who underwent a single-level transforaminal epidural steroid injection was analyzed. Nine input variables were included: age, gender, injection side, injection level, pain duration, pretreatment Numeric Rating Scale pain score, predominant stenotic location, and the degrees of central and foraminal stenosis. A positive therapeutic outcome was defined as a ≥ 50% reduction in the Numeric Rating Scale score at one month postprocedure. Deep learning neural network, XGBoost, and CatBoost algorithms were used to develop the models. Model performance was evaluated using receiver operating characteristic analysis and standard classification metrics.

RESULTS

The deep learning neural network model, configured with 5 fully connected layers, had the best performance among the 3 models, achieving a validation accuracy of 0.868 and an area under the curve (AUC) of 0.882 (95% CI, 0.733-0.994). CatBoost had an intermediate performance with an accuracy of 0.789 and an AUC of 0.765 (95% CI, 0.593-0.906), whereas XGBoost had the lowest accuracy of 0.763 and an AUC of 0.662 (95% CI, 0.458-0.853). In addition, the deep learning neural network model achieved balanced precision and recall across both favorable and poor outcome classes, with a weighted F1 score of 0.869.

LIMITATIONS: This was a single-center retrospective study with a small sample size.

CONCLUSIONS

Deep learning demonstrated superior predictive performance compared with conventional machine learning algorithms in forecasting transforaminal epidural steroid injection outcomes among patients with lumbar spinal stenosis. Integrating clinical and magnetic resonance imaging-derived features through a deep learning neural network may facilitate individualized treatment decisions and enhance therapeutic planning for lumbosacral radiculopathy. Future multicenter studies with larger datasets are warranted for external validation and clinical implementation.