<p>Accurately predicting mortality using medical information is a critical task in the healthcare profession. The common method is extracting features from electronic health records and choosing appropriate models to identify patients who may have a higher risk of death. However, the lack of standardization and identifiability in the storage of unstructured data in medical records poses a significant challenge in training models to predict patient mortality. In this paper, we propose a heterogeneous data fusion mortality prediction model called Fusion-transformer to address the above issues. The proposed Fusion-transformer model uses a time-aware self-attention mechanism to combine structured data represented by time signals and static information and unstructured data represented by clinical notes to enhance patient representation learning. Fusion-transformer also uses a time decay module to control the influence of the patient’s hospitalization experience on the current health status at different times. The results of in-hospital mortality prediction experiments on the Medical Information Mart for Intensive Care III dataset show that Fusion-transformer outperforms other machine learning and deep learning baseline models, improving the area under the receiver operating characteristics curve metric from 0.891 to 0.923. In addition, we explore the practical applications value of our method through real ICU patient electronic medical records.</p>