Concrete-filled steel tubular (CFST) columns are widely used in the construction of civil engineering structures due to their excellent structural performance, which are able to utilize advantages of both concrete and steel materials. The axial compressive strength is considered to be one of the most critical parameters used in the design of CFST columns, and different formulae have been provided in codes and specifications to predict the design strength values. In this paper, 1581 CFST column concentric compressive load test data points are collected from a variety of scientific literature published in China, USA and Japan to establish an extensive database. Random forest (RF) algorithm is adopted to predict the axial compressive strength of CFST columns by using a randomly assigned test set from the database. Two types of features are selected as key input parameters in the development of RF model, including basic parameters and intermediate parameters. In particular, indirect parameters with physical meanings defined in Chinese design codes are used as input features. The comparison of the proposed RF model with the existing code-specified equations underscores the efficiency and accuracy of RF technique in the predication of axial compressive strength. Furthermore, higher prediction accuracy is achieved by machine learning techniques based on intermediate parameters than those based on original parameters.