Abstract: Accurate identification of sepsis patients with higher mortality is of great significance to improve survival outcomes and assist ICU doctors in medical decision-making. However, traditional machine learning methods require complex feature engineering and do not fully utilize the high-missing dynamic time-series data and sparse static data in medical data of patient. Aiming at the shortcomings of mortality prediction of sepsis patients in ICU, an attention-based multiple input fusion learning model was designed to capture patient’s features in the spatiotemporal dimension of patient medical records from dynamic time-series data with high missing rate and sparse static data, respectively, and learn the interaction between spatiotemporal features. The medical records of 10,567 ICU Sepsis patients with the definition of Sepsis-3 were extracted from the MIMIC-Ⅲ dataset. The data were divided into a training set and a test set with the ratio of 8∶2, and the performance of the model was evaluated on the test set using the 5-fold cross-validation on the training set. The experimental results showed that the model had relatively higher AUROC and AUPRC compared with the baseline method, and effectively improved the performance of mortality prediction in ICU sepsis patients.