Machine-Learning-Based Prediction of Cervical Pedicle Screw Malposition from Clinical and Anatomical Features

Area of Science:

• Spine Surgery
• Medical Imaging
• Machine Learning

Background:

• Cervical pedicle screw (CPS) placement offers biomechanical stability but carries risks of malposition.
• Current imaging and navigation improve accuracy but reliable malposition prediction remains difficult.

Purpose of the Study:

• To evaluate machine learning (ML) models for predicting CPS malposition using clinical and anatomical features.
• To assess the interpretability and performance of various ML models in this prediction task.

Main Methods:

• Retrospective analysis of 862 CPS from 168 surgeries.
• Inclusion of clinical, procedural, and anatomical variables (e.g., age, sex, pedicle angle/width).
• Training and comparison of supervised ML models (Random Forest, XGBoost, SVM, KNN) using Python/scikit-learn.

Main Results:

• Moderate and consistent predictive performance across models; Balanced Random Forest showed highest ROC AUC (0.69).
• SHAP analysis identified pedicle width and angle as key predictors.
• Predictive performance was influenced by dataset characteristics like class distribution and feature overlap.

Conclusions:

• ML models can consistently and interpretably predict CPS malposition using available features.
• Dataset characteristics are more critical than model selection for predictive accuracy.
• This study establishes a baseline for ML-based CPS prediction, advocating for larger datasets and detailed anatomical data in future research.