Mapping between EQ-5D-5L and EQ-5D-3L in lumbar spine surgery patients: Machine learning compared to crosswalk algorithms.

Introduction: The transition from EQ-5D-3L to EQ-5D-5L poses challenges for comparing health-related quality of life across studies and clinical settings. Although official EuroQol crosswalk algorithms are commonly used to map between versions, alternative mapping approaches may be relevant for disease-specific populations. Research question: This study evaluated the performance of machine-learning models for bidirectional mapping between EQ-5D-3L and EQ-5D-5L in patients undergoing evaluation for elective lumbar spine surgery for degenerative conditions. Material and methods: Paired EQ-5D-3L and EQ-5D-5L single-center responses (n = 766) were analyzed using repeated cross-validation. Response-level mapping was performed using parametric ordinal regression models (VGAM, ordinalNet) and tree-based machine-learning models (XGBoost, CatBoost, ordinalForest). Both hard and soft prediction approaches were evaluated on held-out data and compared with the official crosswalk algorithms using RMSE, MAE, bias (mean signed error), R Results: Soft predictions generally showed lower prediction errors and higher predictive R Discussion and conclusion: Machine-learning models provided alternative approaches to the official EuroQol crosswalk algorithms, with performance varying by mapping direction and prediction approach. Soft prediction generally showed more favorable predictive performance, particularly for 3L-to-5L mapping. Overall, these findings support further investigation of response-level mapping machine-learning approaches to harmonize mixed-format EQ-5D data.