Traffic accidents have become one of the biggest public health safety matters, which has raised many concerns from citizens and city managers. Accurate traffic accident prediction can not only assist the government in making decisions in advance but also enhance public trust in public safety. Conventional spatio-temporal prediction models, limited by the skewed distributions and sparse labels of traffic accident occurrence, are prone to overfitting. Inspired by hypergraph learning and self-supervised learning, this paper suggests a sparse spatio-temporal dynamic hypergraph learning (SST-DHL) framework to capture the higher-order dependencies in sparse traffic accidents. Specifically, a multi-view spatio-temporal convolution block is employed first to capture the local spatio-temporal correlation and inherent semantics of traffic accidents. Then we propose a cross-regional dynamic hypergraph learning model to capture global spatio-temporal dependencies beneath the entire urban landscape. In addition, a two-supervised self-learning paradigm is intended to strengthen the representation of sparse traffic occurrences by regional self-identification, which can capture local and global spatio-temporal traffic patterns. The proposed model is applicable to most sparse datasets for traffic forecasts. Extensive experiments was conducted on two heterogeneous accident datasets from New York City and London, and the results shows an average improvements of 7.21%-23.09% at different sparsity levels compared to the optimal baselines. More importantly, the proposed SST-DHL improves the interpretability of model results, which demonstrates that hypergraph learning can efficiently capture the complex higher-order spatio-temporal dependencies among different traffic accident instances.