Deviations in alternative splicing are increasingly implicated in cancer progression, generating oncogenic protein variants that drive tumorigenesis. However, the molecular mechanisms underlying alternative splicing's role in cancer prognosis remain unclear. Here, we employ deep learning to explore the impact of alternative splicing events on cancer type and prognosis. Our prediction models for each alternative splicing type across 26 cancer types from over 9,000 samples reveals that alternative splicing exhibits comparable predictive power to gene expression in cancer type classification. Remarkably, our analysis of 234 prediction models shows alternative splicing outperforms gene expression in predicting patient survival time across 19 cancer types. Building a multimodal model composed of fusing separate pre-tuned encoders using a gated approach with an MLP and cox regression head, we find the same heterogenous result in terms of best performer. Of significant note, we find that combining the information in the multimodal model boost performance in predictive power of concordance index. Our findings underscore the superior predictive power of alternative splicing over traditional gene expression analysis and highlight novel therapeutic opportunities for targeted cancer treatments.