Abstract: Predicting mortality of ICU patients with high accuracy is an active research in clinical domain during the past decades. However, the special features of ICU data such as high-dimensional, uncertain sampling and imbalanced distribution makes the prediction challenging. In this study, a hierarchical data model is proposed to describe the special feature of ICU data. A hybrid framework with clustering and machine learning algorithm is used to convert the ICU time series with special data property to the traditional time series data so that time series data mining can be applied. High dimension of ICU data is reduced by time series clustering while, uncertain sampling is settled by certainty strategy. Additionally, the unknown medical knowledge exists both within variables and among variables which is difficult to extract. To address this issue, proposed framework used clustering to extract knowledge within the variable and knowledge among variables is extracted by classical machine learning algorithms. Experimental results show prediction accuracy of proposed hybrid framework is better than the data mining methods which do not consider analysis of ICU data properties and additionally, more efficient results can be achieved with suitable choice of sampling frequency.