Super enhancers (SEs) are critical cis-regulatory elements that govern cell identity and transcriptional activity, playing essential roles in various diseases. Traditional super enhancers prediction methods rely on ChIP-seq and DNase-seq data, which often contain numerous irrelevant features and fail to integrate multiple data modalities effectively, thereby reducing the predictive performance. To address these issues, we propose a novel deep learning model, DEEPFUSION, which predicts SEs by integrating DNA sequence data with epigenomic features. This model employs causal inference to identify the most relevant features while accounting for confounding variables. By utilizing dna2vec, the model encodes DNA sequences into feature representations and then integrates them with epigenomic data, through advanced feature fusion technology. The experimental results demonstrate that the features selected through causal inference can identify key features relevant to splicing event prediction and reduce computational costs. Moreover, multimodal data integration further improves the interpretability and robust ness of the predictive model. Overall, DEEPFUSION achieves superior performance compared to existing models, underscoring its potential to advance bioinformatics research.