# Machine Learning Prediction of Therapeutic Outcome After Transforaminal Epidural Steroid Injection for Radiculopathy from Herniated Lumbar Disc

## TL;DR

This study uses machine learning to predict the success of a spinal injection treatment for lower back pain, showing that a deep learning model performs best.

## Contribution

The first ML model integrating clinical and MRI data to predict outcomes of TFESI for HLD-related radiculopathy.

## Key findings

- A deep neural network model achieved the highest predictive accuracy (AUC 0.821) for TFESI outcomes.
- The DNN model outperformed random forest and XGBoost in distinguishing favorable from poor outcomes.
- Integration of clinical and MRI data improved outcome prediction for spinal injections.

## Abstract

Background/Objectives: Transforaminal epidural steroid injection (TFESI) is widely used to treat lumbosacral radicular pain caused by a herniated lumbar disc (HLD). However, therapeutic response varies substantially, and reliable outcome prediction remains challenging because of the multifactorial interplay of clinical and morphological factors. Machine learning (ML) approaches may address this limitation by modeling nonlinear interactions among patient-specific variables. Methods: This retrospective cohort study analyzed 242 patients with HLD-related radiculopathy who underwent single-level lumbar TFESI. Eight variables—age, sex, injection side, injection level, pain duration, pretreatment numeric rating scale (NRS) score, HLD location, and HLD subtype—were used as input features. Therapeutic outcome was defined as a ≥50% reduction in NRS score at 1 month after TFESI. Three predictive models, namely deep neural network (DNN), random forest (RF), and XGBoost, were developed and evaluated using a validation cohort of 49 patients. Results: The DNN model demonstrated the best validation performance, achieving an area under the curve (AUC) of 0.821 (95% confidence interval [CI], 0.690–0.929). The performance of the RF (AUC, 0.711; 95% CI, 0.535–0.865) and XGBoost (AUC, 0.674; 95% CI, 0.498–0.831) models was inferior to that of the DNN. In addition, the DNN produced fewer false-positive predictions and showed more robust discrimination between favorable and poor outcomes than the other ML models. Conclusions: A deep learning–based predictive model demonstrated superior performance in predicting therapeutic outcomes after lumbar TFESI in patients with HLD-related radiculopathy. Integration of routine clinical and magnetic resonance imaging (MRI)-derived features into ML algorithms may enhance individualized prognostication and assist clinicians in optimizing patient selection for interventional procedures. To the best of our knowledge, this is the first study to develop an ML-based model integrating routine clinical variables with MRI findings for the prediction of TFESI outcomes in HLD-related radiculopathy. Nevertheless, the study is limited by its single-center retrospective design, lack of external validation, and reliance on MRI assessments performed by a single rater. Future multicenter studies are warranted to improve generalizability and confirm clinical utility.