Predicting postoperative neurological outcomes of degenerative cervical myelopathy based on machine learning.

Abstract

Introduction: This study aimed to develop machine learning models to predict neurological outcomes in patients with degenerative cervical myelopathy (DCM) after surgical decompression and identify key factors that contribute to a better outcome, providing a reference for patient consultation and surgical decision-making.

Methods: This retrospective study reviewed 1,895 patients who underwent cervical decompression surgery for DCM at Peking University Third Hospital from 2011 to 2020, with 672 patients included in the final analysis. Five machine learning methods, namely, linear regression (LR), support vector machines (SVM), random forest (RF), XGBoost, and Light Gradient Boosting Machine (LightGBM), were used to predict whether patients achieved the minimal clinically important difference (MCID) in the improvement in the Japanese Orthopedic Association (JOA) score, which was based on basic information, symptoms, physical examination signs, intramedullary high signals on T2-weighted (T2WI) magnetic resonance imaging (MRI), and various scale scores. After training and optimizing multiple ML algorithms, we generated a model with the highest area under the receiver operating characteristic curve (AUROC) to predict short-term outcomes following DCM surgery. We evaluated the importance of the features and created a feature-reduced model. The model's performance was assessed using an external dataset.

Results: The LightGBM algorithm performed the best in predicting short-term neurological outcomes in the testing dataset, achieving an AUROC value of 0.745 and an area under the precision-recall curve (AUPRC) value of 0.810. The important features influencing performance in the short-term model included the preoperative JOA score, age, SF-36-GH, SF-36-BP, and SF-36-PF. The feature-reduced LightGBM model, which achieved an AUROC value of 0.734, also showed favorable performance. Moreover, the feature-reduced model showed an AUROC value of 0.785 for predicting the MCID of postoperative JOA in the external dataset, which included 58 patients from other hospitals.

Conclusion: We developed models based on machine learning to predict postoperative neurological outcomes. The LightGBM model presented the best predictive power regarding the surgical outcomes of DCM patients. Feature importance analysis revealed that variables, including age, preoperative JOA score, SF-36-PF, SF-36-GH, and SF-36-BP, were essential factors in the model. The feature-reduced LightGBM model, designed for ease of application, achieved nearly the same predictive power with fewer variables.