The massive amount of Electronic Health Records provides a significant opportunity for improving clinical decision-making. Prevalent multimodality methods focus on integrating multimodality to get a unified patient representation for medical outcome tasks, neglecting the computation efficiency of the time-series modality encoder and the prediction reliability. To tackle these limitations, this study proposes an efficient time-series representation with an evidential multimodal fusion framework for reliable mortality risk prediction. First, the modality-specific representation learning extracts the representations from time-series, demographic, and clinical note modalities. Specifically, we propose to encode the time-series data using an efficient parallel attention encoder, which tackles the limitation of the computation complexity of the Transformer and avoids the gradient vanishing problem of the RNNs. Second, the evidential multimodal fusion module is proposed to adaptively assign weights to highly reliable modalities based on their uncertainty, providing a reliable prediction. Specifically, the representation of each modality is converted to evidence and opinions, including the classification probabilities and the uncertainty estimates, for each modality. After that, the evidential fusion module integrates them into a unified, confident representation. The proposed framework enhances the mortality prediction on the MIMIC-III dataset, outperforming several state-of-the-art methods.