To get a highly reliable software product to the market on schedule, software engineers must allocate resources on the fault-prone software modules across the development effort. Software quality models based upon data mining from past projects can identify fault-prone modules in current similar development efforts. So that resources can be focused on fault-prone modules to improve quality prior to release. Many researchers have applied the neural networks approach to predict software quality. Although neural networks have shown their strengths in solving complex problems, their shortcoming of being ?black boxes? models has prevented them from being accepted as a common practice for fault-prone software modules prediction. That is a significant weakness, for without the ability to produce comprehensible decisions, it is hard to trust the reliability of neural networks that address real-world problems. In this paper, we introduce an interpretable neural network model for software quality prediction. First, a three-layer feed-forward neural network with the sigmoid function in hidden units and the identity function in output unit was trained. The data used to train the neural network is collected from an earlier release of a telecommunications software system. Then use clustering genetic algorithm (CGA) to extract comprehensible rules from the trained neural network. We use the rule set extracted from the trained neural network to detect the fault-prone software modules of the later release and compare the predicting results with the neural network predicting results. The comparison shows that although the rule set?s predicting accuracy is a little less than the trained neural network, it is more comprehensible.