LASSO-based machine learning algorithm for prediction of dysphagia in patients suffering anterior cervical discectomy and fusion.

PURPOSE: To develop and validate a predictive model for postoperative dysphagia in patients undergoing anterior cervical discectomy and fusion (ACDF).

METHODS

We retrospectively analyzed 500 patients who underwent ACDF at our institution between 2018 and 2022. A total of 53 candidate predictors-including 21 radiographic measurements, 19 preoperative blood biomarkers, 9 medical history variables, and 4 surgical characteristics-were evaluated. Patients were randomly partitioned into a training cohort (75%, n = 375) and a validation cohort (25%, n = 125). We determine independent prognostic factors used univariate comparisons between dysphagia and non-dysphagia groups, the least absolute shrinkage and selection operator (LASSO) regression, receiver operating characteristic (ROC) curves analysis, followed by an interaction analysis. The LASSO regression model was applied for predictive signature building in the training set. Model performance was assessed by receiver operating characteristic (ROC) curve analysis, decision curve analysis (DCA), and calibration plots in both cohorts.

RESULTS

Within one month after ACDF, 115 of 500 patients (23.0%) developed dysphagia. Univariate analysis yielded 37 variables associated with dysphagia. LASSO regression retained 13 predictors, and ROC analysis identified 30 variables with area under the curve (AUC) > 0.6. Intersecting these sets produced a final panel of 9 key predictors. The model achieved an AUC of 0.969 in the training cohort and 0.954 in the validation cohort. The DCA and calibration curve analysis showed good performance for the diagnostic model in both sets.

CONCLUSION

A LASSO-penalized logistic regression model incorporating radiographic parameters, blood biomarkers, medical history variables, and surgical characteristics accurately predicts postoperative dysphagia following ACDF and may guide individualized risk stratification.