Using K-nearest neighbor (KNN) to predict mortality is an important mean to positively affect patient health. However, it is difficult for KNN to use a single distance to accurately measure the distance of samples with discrete and continuous variables. Furthermore, the voting method applied in KNN cannot measure the impact of distance on results. To solve above problems, a KNN mortality prediction model with mixed weighted distance was proposed. First, the chi-square test and logistic regression with L1 regularization are used for feature selection and ranking. Next, a mixture of Value Difference Metric (VDM) and Manhattan distance is applied to calculate the distance. Then, the softmin function is chosen to weight the distance and finally give the category for testing sample. In the end, the data of 2 743 htart failure patients in the MIMIC-III public database were experimentally evaluated, which verifies that the improved algorithm has a good performance in mortality prediction.