sparse Proper Generalized Decomposition (s-PGD) is a powerful data-driven regression technique for approximating high-dimensional response functions, thereby enabling efficient uncertainty quantification and design under uncertainty [1]. However, s-PGD may still encounter difficulties in high-dimensional nonlinear problems, where enforcing a low-rank separated structure becomes challenging. This work proposes a novel surrogate modeling method that integrates s-PGD with an ANOVA-based high-dimensional model representation (ANOVA-HDMR) for uncertainty quantification of strongly nonlinear responses with high-dimensional inputs [2]. The proposed method preserves the main advantages of s-PGD, including its scalability and flexibility, while explicitly adhering to the ANOVA-HDMR decomposition structure. This efficiently approximates statistical moments, such as the mean and variance, without explicitly constructing high-order interaction terms in the basis modes. We demonstrate the accuracy and efficiency of the proposed method using a 20-dimensional numerical example that exhibits nonlinear behavior. We conduct a comparative study against a polynomial chaos expansion (PCE) surrogate under identical sampling conditions. The results show that the proposed ANOVA-based s-PGD method predicts statistical moments more accurately, highlighting its robustness and efficiency in high-dimensional settings. The proposed method provides a scalable and practical alternative to conventional ANOVA-based surrogate models for engineering applications involving high dimensionality.