With the rapid growth of observational data in various scientific and technological domains, such as healthcare and power systems, causal inference research has gained significant momentum. However, estimating causal effects using observational data faces two primary obstacles: the fundamental absence of counterfactual outcomes and the presence of confounding factors. These challenges become more pronounced when dealing with high-dimensional covariates, as existing methods often fail to capture their inherent correlations. To address these issues, this paper proposes a Causal Encoding Generative Model based on the Attention mechanism and the Kolmogorov-Arnold Network (KAN), referred to as CAKEGM. The model first applies an attention mechanism to learn the internal dependencies among covariates, then performs dimensionality reduction to map features into a low-dimensional latent space with known density. Simultaneously, KAN is employed to fit treatment variables and outcomes, enabling effective disentanglement of covariate representations in the latent space. We evaluate CAKEGM on two healthcare-related datasets. Experimental results show that CAKEGM outperforms existing state-of-the-art methods in both accuracy and robustness for treatment effect estimation.