Mapping FACT-Hep to SF-6Dv2 utility values in patients with primary liver cancer: a comparison of conventional regression and machine learning approaches

Abstract Purpose To develop and internally validate mapping algorithms for predicting Short Form Six-Dimension version 2 (SF-6Dv2) utility values from the Functional Assessment of Cancer Therapy-Hepatobiliary (FACT-Hep) in patients with primary liver cancer, and to compare conventional regression and machine-learning approaches. Methodology A multicenter cross-sectional study included 608 patients from two hospitals in Sichuan Province, China. FACT-Hep, SF-6Dv2, demographic, and clinical data were collected. Conceptual overlap and multicollinearity were examined before model development. Five predictor specifications were evaluated using ordinary least squares (OLS), Tobit, ordered probit (Oprobit), ridge regression, support vector regression (SVR), and gradient boosting regression (GBR). The two best specifications from each method were retained for fivefold cross-validation and random 70/30 holdout validation. Model selection was based on prediction-error and agreement measures, average rank value (ARV), and graphical assessment. Results The mean observed SF-6Dv2 utility was 0.7332 (SD 0.2306), ranging from − 0.2270 to 1.0000. Oprobit5 performed best among the conventional models in the initial analysis, whereas GBR5 showed the strongest apparent performance among machine-learning models. After integrating fivefold cross-validation and random holdout validation, GBR4 achieved the lowest integrated ARV (3.05), followed by GBR5 (3.50), SVR4 (4.00), Oprobit5 (4.15), and SVR3 (4.25). Graphical analyses showed closer agreement in the central utility range but regression toward the mean at the extremes. Conclusion FACT-Hep can be mapped to SF-6Dv2 utility values with acceptable predictive performance. GBR4 showed the best overall performance after integrating the two internal validation procedures, while Oprobit5 was the strongest conventional alternative. Independent external validation is needed before wider application.