Development and Validation of an MRI-Based Radiomics Model Integrating Spinal Canal and Paraspinal Muscle Features for Predicting Early Functional Recovery After Unilateral Biportal Endoscopic Decompression in Lumbar Spinal Stenosis

<title>Abstract</title> <p>Purpose To develop and validate machine learning models based on single-sequence T2-weighted fat-suppressed (T2-FS) magnetic resonance imaging (MRI) radiomic features extracted from the spinal canal (SC) and the paravertebral muscle (PVM) compartment for predicting early functional recovery following single-level unilateral biportal endoscopic decompression in patients with lumbar spinal stenosis. Methods A total of 175 patients with LSS who underwent single-level UBE decompression between January 2024 and April 2026 were retrospectively enrolled. The primary endpoint was achievement of the minimum clinically important difference (MCID) in the Oswestry Disability Index (ODI) at two months postoperatively, defined as an absolute improvement of at least 12.8 points, or a relative improvement of at least 50% for patients with a preoperative ODI of 26 or lower. Radiomic features were extracted from preoperative T2-FS axial images using PyRadiomics within manually segmented SC and PVM regions of interest (ROI). Three feature sets were constructed, namely SC alone, PVM alone, and Combined set. After sequential feature selection via Spearman correlation analysis, recursive feature elimination (RFE), and the least absolute shrinkage and selection operator (LASSO), seven machine learning classifiers were trained and compared. The model’s performance was evaluated using the area under the curve (AUC) metric. Results Among the 175 patients, 130 (74.3%) achieved MCID and 45 (25.7%) did not. After feature selection, six radiomic features were retained for the SC model, six for the PVM model, and eight for the Combined model. In the independent test cohort, the Combined model using Gaussian Naive Bayes (GNB) yielded the highest AUC of 0.897 (95% CI: 0.744–1.000), outperforming the best SC model (Random Forest; AUC 0.861, 95% CI: 0.720–0.984) and the best PVM model (Logistic Regression; AUC 0.710, 95% CI: 0.535–0.865). Calibration curves and decision curve analysis (DCA) confirmed the Combined model’s favorable calibration and the greatest net clinical benefit across a broad range of threshold probabilities. Conclusion A radiomics model combining SC and PVM features derived from a single T2-FS MRI sequence predicts early functional recovery after Unilateral biportal endoscopic decompression with good discriminatory performance, surpassing models built on either region alone. This noninvasive imaging tool holds promise for preoperative risk stratification in patients undergoing minimally invasive lumbar spine surgery.</p>