Federated Learning (FL) is proposed to overcome data silos in model training of Machine Learning (ML) through joint modeling with privacy-preserving. Mobile Network Operators (MNOs) and Healthcare Providers (HPs) share common users with different features. Data silo exists due to the inaccessibility from MNOs to HPs for privacy-preserving. Therefore, the data of MNOs and HPs can be utilized together in FL settings to empower each other's service. In this paper, we established a Federated Gradient Boosting Decision Tree (FGBDT) system based on SecureBoost, a lossless federated ensemble model found on GBDT, to improve user classification for HPs in Healthcare Recommendation with MNOs' data. Our experiment shows that FGBDT has a 9.71% of precision increase and a 4% of F1 score improvement, and a 10.45% of cumulative precision improvement in a real operational practice than GBDT.