MRI machine learning model predicts nerve root sedimentation in lumbar stenosis: a prospective study.

OBJECTIVES

To analyze MRI characteristics of the nerve root sedimentation sign (SedSign) in lumbar spinal canal stenosis (LSS) and to establish a risk model predicting its occurrence.

METHODS

A total of 1,138 narrow layers were divided into SedSign-positive (426 layers) and SedSign-negative (712 layers) groups. Key data included spinal canal diameters, dural sac dimensions, ligamentum flavum (LF) and epidural fat (EF) thickness, SedSign presence, lumbar disc herniation (LDH), high-intensity zone (HIZ), and EF classification. Comparisons used t tests or Mann-Whitney U tests. Recursive feature elimination with cross-validation (RFECV) was used to select predictive features, and models were established via random forest (RF), K-nearest neighbors (KNN), and extreme gradient boosting (XGBoost) algorithms and evaluated in terms of precision, recall, average F1 score, accuracy, and AUC. The optimal model was subject to SHAP analysis to explain the risk factors.

RESULTS

LSS patients with the SedSign had a greater degree of narrowing and were more likely to have increased EF, LDH, LF hypertrophy (LFH), and HIZ and to be older than those without the SedSign. There was no difference between the two groups in terms of sex ( = 0.051). RFECV yielded eight features: age, sex, APDS, APDD, TDD, EF grade, LDH, and LFH. The RF model constructed using these features-designated as SedSign8-exhibited superior performance in predicting the risk of SedSign, with robust metrics across all evaluation dimensions: precision of 84.4%, recall of 73.6%, F1 score of 78.6%, accuracy of 83.6%, and an AUC of 0.901.

CONCLUSION

Older patients, along with a greater degree of stenosis and changes in the dural sac and surrounding tissue structures, were identified as the main pathophysiological basis for the occurrence of the SedSign in LSS.