Exploring the Potential of Machine Learning for Mapping FACT‑L to EQ‑5D‑5L in Lung Cancer

Background: To evaluate machine learning approaches for mapping FACT-L to EQ-5D-5L utilities in lung cancer patients, with a particular focus on exploring neural network models and their predictive performance. Methods: We enrolled 347 hospitalized lung cancer patients at West China Tianfu Hospital of Sichuan University and collected sociodemographic, clinical, FACT-L, and EQ-5D-5L data. Random forest (RF), extreme gradient boosting (XGBoost), support vector regression (SVR), and neural networks (NN) were evaluated using five-fold cross-validation. Performance was assessed with mean squared error (MSE) and mean absolute error (MAE), and we examined input dimensionality and neural network architecture. Results: Using all FACT-L items, SVR achieved the best performance (MSE ≈ 0.033, MAE ≈ 0.136), with RF and XGBoost close behind. Model performance improved with increasing input dimensionality, with the best results obtained using all FACT-L items. Among neural networks, the simple architecture (one hidden layer, 32 neurons) achieved competitive performance. Conclusion: Machine learning improves FACT-L to EQ-5D-5L mapping predictive performance; SVR performed best under small-to-medium dataset conditions, and appropriately designed neural networks show promise for utility estimation in lung cancer.