Development and Validation of a Machine Learning-Based Online Prognostic Model for Cervical Spondylosis Patients After Anterior Cervical Discectomy and Fusion: A Multicenter Study

Abstract

Background

Cervical spondylosis (CS) is a degenerative condition often requiring surgical intervention, such as anterior cervical discectomy and fusion (ACDF), to alleviate symptoms. However, postoperative outcomes can vary significantly. This study aimed to develop and validate a predictive model for 1-year outcomes in CS patients after ACDF using multiple machine learning algorithms.

Methods

Data from 973 patients across three clinical centers, including 872 patients in the retrospective cohort and 101 patients in the prospective cohort, were utilized. A variety of clinical and laboratory features were identified using LASSO regression. Various machine learning algorithms were employed to develop predictive models. The models' performance was assessed and compared using metrics such as receiver operating characteristic (ROC) curves, area under the curve (AUC), calibration analysis, and decision curve analysis (DCA). Model interpretation and feature importance analysis were carried out using the SHapley Additive exPlanations (SHAP) method. Finally, the model was deployed on the web by using the Shiny app.

Results

The model was constructed using 10 essential predictors. Ten machine learning models were evaluated, with the stacking ensemble learning model demonstrating superior predictive performance (AUC = 0.81 in the internal validation set, 0.80 in the external validation set, and 0.82 in the prospective cohort). Furthermore, CRP, MONO, ESR, and age were highlighted as critical predictors.

Conclusions

This predictive tool offers a robust framework for personalized postoperative management in CS patients, potentially improving clinical outcomes.