Machine-learning models for the prediction of ideal surgical outcomes in patients with adult spinal deformity.

Abstract

Aims: Adult spinal deformity (ASD) surgery can reduce pain and disability. However, the actual surgical efficacy of ASD in doing so is far from desirable, with frequent complications and limited improvement in quality of life. The accurate prediction of surgical outcome is crucial to the process of clinical decision-making. Consequently, the aim of this study was to develop and validate a model for predicting an ideal surgical outcome (ISO) two years after ASD surgery.

Methods: We conducted a retrospective analysis of 458 consecutive patients who had undergone spinal fusion surgery for ASD between January 2016 and June 2022. The outcome of interest was achievement of the ISO, defined as an improvement in patient-reported outcomes exceeding the minimal clinically important difference, with no postoperative complications. Three machine-learning (ML) algorithms - LASSO, RFE, and Boruta - were used to identify key variables from the collected data. The dataset was randomly split into training (60%) and test (40%) sets. Five different ML models were trained, including logistic regression, random forest, XGBoost, LightGBM, and multilayer perceptron. The primary model evaluation metric was area under the receiver operating characteristic curve (AUROC).

Results: The analysis included 208 patients (mean age 64.62 years (SD 8.21); 48 male (23.1%), 160 female (76.9%)). Overall, 42.8% of patients (89/208) achieved the ideal surgical outcome. Eight features were identified as key variables affecting prognosis: depression, osteoporosis, frailty, failure of pelvic compensation, relative functional cross-sectional area of the paraspinal muscles, postoperative sacral slope, pelvic tilt match, and sagittal age-adjusted score match. The best prediction model was LightGBM, achieving the following performance metrics: AUROC 0.888 (95% CI 0.810 to 0.966); accuracy 0.843; sensitivity 0.829; specificity 0.854; positive predictive value 0.806; and negative predictive value 0.872.

Conclusion: In this prognostic study, we developed a machine-learning model that accurately predicted outcome after surgery for ASD. The model is built on routinely modifiable indicators, thereby facilitating its integration into clinical practice to promote optimized decision-making.