A stacking ensemble machine learning model for predicting postoperative axial pain intensity in patients with degenerative cervical myelopathy

Machine learning (ML) has been extensively utilized to predict complications associated with various diseases. This study aimed to develop ML-based classifiers employing a stacking ensemble strategy to forecast the intensity of postoperative axial pain (PAP) in patients diagnosed with degenerative cervical myelopathy (DCM). A total of 711 consecutive postoperative DCM patients were included between 2016 and 2024, and after excluding patients who did not meet the inclusion criteria and those who met the exclusion criteria, a total of 484 patients were ultimately included in this study. The intensity of PAP was assessed using a standardized Numerical Rating Scale (NRS) score one year following surgery. Participants were randomly allocated into training and testing sub-datasets in a ratio of 8:2. 91 initial ML classifiers were developed, from which the top three highest-performing classifiers were subsequently integrated into an ensemble model utilizing 13 different machine learning models. The area under the curve (AUC) served as the primary metric for evaluating the predictive performance of all classifiers. The classifiers EmbeddingLR-RF (AUC = 0.81), EmbeddingRF-MLP (AUC = 0.81), and RFE-SVM (AUC = 0.80) were recognized as the leading three models. By implementing an ensemble learning approach such as stacking, an enhancement in performance for the ML classifier was observed after amalgamating these three models, with SVM ensemble classifier performed the best (AUC = 0.91). Decision curve analysis underscored the advantages conferred by these ensemble classifiers; notably, prediction curves for PAP intensity among DCM patients exhibited significant variability across the top three initial classifiers. The ensemble classifiers effectively predicted PAP intensity in DCM patients, showcasing substantial potential to aid clinicians in managing DCM cases while optimizing medical resource utilization.