401 How do Classification Versus Regression Machine Learning Algorithms Predict 5-Year Minimum Clinically Important Differences in QOD Lumbar Spondylolisthesis Patients?

INTRODUCTION: Machine learning techniques able to parse high-dimensional clinical data are a promising avenue for optimized identification of ideal surgical candidates. The formulation of such medical prediction problems as classification or regression tasks can have significant effects on model performance. METHODS: Patients undergoing single-segment surgery were collected from the prospective Quality Outcomes Database Lumbar Spondylolisthesis dataset. Classification models were trained on 60-month achievement of minimum clinically important differences (MCID) for ODI, NRS-Back Pain, and NRS-Leg Pain. Equivalent regression models were trained on continuous variables of 60-month change in ODI, NRS-BP, and NRS-LP. Predicted values were converted to MCID achievement for model evaluation. All models were trained on an 80/20 split with 5-fold cross-validation, and Bayesian search was used for hyperparameter tuning. RESULTS: Our final cohort included 608 patients, with 487 completing 60-month follow-up. Of these patients, 304 (62.9%) reached MCID for ODI, 280 (66.2%) reached MCID for NRS-BP, and 322 (69.7%) reached MCID for NRS-LP. All models demonstrated highest predictive performance on MCID for NRS-LP tasks. The classification and regression models demonstrated largely equivalent AUC across prediction of MCID for NRS-LP (C:0.70±0.07to0.76±0.05, R:0.66±0.05to0.77±0.05), NRS-BP (C:0.63±0.06to0.72±0.07, R:0.67±0.07to0.69±0.07), and ODI (C:0.60±0.04to0.67±0.04, R:0.50±0.00to0.64±0.05). Notably, the MLP regressor outperformed all models on AUC for NRS-LP MCID prediction. For NRS-LP specifically, regression models consistently demonstrated higher accuracy, precision, F1, and recall metrics versus their classification counterparts. CONCLUSIONS: In a cohort undergoing surgery for Grade 1 degenerative lumbar spondylolisthesis, ML algorithms reliably predicted 5-year achievement of MCID across PROs. The equivalent and sometimes superior performance of regression models emphasizes the importance of thoughtful model selection and task formulation guided by clinical context.