Machine learning models for predicting patient satisfaction after adult spinal deformity surgery

OBJECTIVE: Patient satisfaction serves as a valuable measure for evaluating outcomes from the patient's perspective. However, the factors critical for predicting satisfaction in patients with adult spinal deformity (ASD) remain elusive. This study aimed to develop and validate predictive models for assessing patient satisfaction 24 months after ASD surgery. METHODS: A total of 213 individuals diagnosed with ASD met inclusion criteria; 128 (60%) patients were randomly selected for model development (training set), and the remaining 85 (40%) were used for internal validation (test set) to assess model robustness. The primary outcome was the satisfaction score from Scoliosis Research Society-22r domains, with scores ≥ 4.5 indicating high satisfaction. Three machine learning (ML) algorithms (least absolute shrinkage and selection operator, recursive feature elimination, and Boruta) were used to identify critical variables for patient satisfaction. A logistic regression model was developed and tested based on these variables to predict personalized satisfaction. Feature importance was ranked using the Shapley Additive Explanations (SHAP) method. RESULTS: ML algorithms identified 9 key indicators of postoperative satisfaction. The predictive model demonstrated an area under the receiver operating characteristic curve of 0.846 and an accuracy of 0.812 in the test set. SHAP analysis revealed that predictors such as improved postoperative Western Ontario and McMaster Universities Osteoarthritis Index (WOMAC) function, absence of frailty, achievement of minimum clinically important difference in imaging, and enhanced WOMAC function increased the relative functional cross-sectional area, higher postoperative subtotal scores, smaller postoperative sagittal vertical axis, successful pelvic compensation, and reduced fatty infiltration significantly influenced postoperative satisfaction. CONCLUSIONS: The results of this study suggest that the developed models can provide patients with personalized prognostic information. Surgeons should consider these routinely modifiable indicators in clinical practice to guide postoperative rehabilitation.