With the development of biomedical equipment and healthcare level, large amounts of data have been brought out in hospital, especially in Intensive Care Unit (ICU). However, how to better exploit meaningful information from these rich data still remains a challenge. This paper focuses on ICU mortality prediction, which is a typical example of second use of ICU big data. Patient ICU mortality prediction faces challenges in many aspects, such as high dimensionality, imbalance distribution and time asynchronization etc. To solve these challenges, a series of analytical methods and tools, including variables selection, preprocessing, feature extraction & feature selection and predictive modeling, have been utilized and developed. High-dimensional and unbalanced natures of the ICU data badly affect the performance of classifiers. We modified the cost-sensitive principal component analysis (CSPCA), which is denoted by MCSPCA, to handle these problems in feature extraction stage. As for parameter optimization, a variant of standard particle swarm optimization called chaos particle swarm optimization (CPSO) was adopted for its capacity of finding optimal solution. In order to obtain the best prediction model, different algorithms were investigated and their AUC performances were evaluated in a large real world benchmark data. The final results show that our proposed method improved the performance of the traditional machine learning methods, in which the support vector machine (SVM) reach best AUC performance of 0.7718. This study gives a paradigm to handle similar problems in big health data and helps promote healthcare services. (C) 2018 Published by Elsevier B.V.