Development and Validation of Interpretable Machine Learning Models Incorporating Paraspinal Muscle Quality to Predict Cage Subsidence Risk Following Posterior Lumbar Interbody Fusion.

Study design: A real-world, multicenter retrospective study.

Objective: To identify independent risk factors for cage subsidence following Posterior Lumbar Interbody Fusion (PLIF) and develop an interpretable machine learning model for risk prediction.

Materials and methods: Patients with degenerative lumbar disease who underwent single-level PLIF (January 2018-October 2023) were retrospectively included. A training set (n=620) came from the First Affiliated Hospital of Soochow University, and a validation set (n=100) from the Second Affiliated Hospital. Cage subsidence (≥2 mm intervertebral height loss) was assessed radiographically. Parameters included paraspinal muscle indices (Psoas Muscle Index [PMI], Multifidus Muscle Index [MM]), Fat Infiltration [FI]), bone density markers (Hounsfield Unit [HU] value, Vertebral Bone Quality [VBQ], Endplate Bone Quality [EBQ]), cage position, and postoperative alignment. Multivariate logistic regression identified risk factors; multiple machine learning models were developed and evaluated. A web-based tool was created for clinical deployment.

Results: Multivariate analysis identified PMI, FI, HU value, VBQ, cage position, cage height, postoperative Intervertebral Height (IH), corrected IH, and corrected SA as independent risk factors for cage subsidence. Light Gradient Boosting Machine (LightGBM) outperformed other models, achieving the highest AUC (0.9752), accuracy (0.92), and F1-score (0.9216), with the lowest Brier score (0.0660). After excluding indicators related to paravertebral muscle function from the prediction model, the predictive accuracy of the model decreased substantially. (SHapley Additive exPlanations) SHAP analysis confirmed VBQ, PMI, BMI, and EBQ as the most influential predictors. The final model was deployed as a web-based tool for real-time clinical risk assessment.

Conclusions: Key risk factors for PLIF cage subsidence were identified, and a validated machine learning model was developed. The high-performance LightGBM model, deployed in a user-friendly web application, enables spine surgeons to optimize surgical planning and reduce subsidence risk.