Machine learning prediction of surgical approach and clinical heterogeneity in cervical spondylotic myelopathy

Preoperative planning of the surgical approach for cervical spondylotic myelopathy (CSM) is central to precision medicine. This study used supervised machine learning (SML) to build a predictive model for surgical approach selection in CSM and applied unsupervised machine learning (UML) to explore clinical heterogeneity.

In this retrospective study, a development cohort of 884 surgically treated patients with CSM was used for model development after exclusion of patients with incomplete perioperative data. Variables showing significant between-group differences were entered into four supervised machine-learning algorithms, including support vector machine (SVM), extreme gradient boosting (XGBoost), random forest (RF), and generalized linear model (GLM), for feature screening and model development. In addition, an exploratory unsupervised clustering analysis was performed in patients undergoing anterior surgery on the basis of perioperative profile variables to characterize clinical heterogeneity. An independent external validation cohort of 200 patients was further used to evaluate the discrimination, calibration, and clinical utility of the final model.

The final prediction model for surgical approach incorporated four key variables: American Spinal Injury Association impairment scale (ASIA), eosinophils (EOS), total protein (TP), and albumin (Alb), with an area under the curve (AUC) of 0.811 in the development cohort. Model scores were significantly correlated with perioperative indicators. UML identified two exploratory subgroups among patients receiving anterior surgery, indicating distinct perioperative profiles. In the external validation cohort, the model achieved an AUC of 0.726 (95% CI 0.615-0.828), with acceptable calibration and potential clinical net benefit on decision curve analysis.

ASIA, EOS, TP, and Alb were identified as variables associated with surgical approach selection in patients with CSM. The internally derived model showed promising discriminative ability and retained moderate performance in an independent external validation cohort. In addition, exploratory unsupervised clustering suggested the presence of distinct perioperative profiles among patients undergoing anterior surgery. These findings provide preliminary support for the potential transportability of the model, although further prospective multicenter validation remains necessary.