The majority of software faults are present in small number of modules, therefore accurate prediction of fault-prone modules helps improve software quality by focusing testing efforts on a subset of modules. This paper evaluates the use of the faults-slip-through (FST) metric as a potential predictor of fault-prone modules. Rather than predicting the fault-prone modules for the complete test phase, the prediction is done at the specific test levels of integration and system test. We applied eight classification techniques to the task of identifying fault-prone modules, representing a variety of approaches, including a standard statistical technique for classification (logistic regression), tree-structured classifiers (C4.5 and random forests), a Bayesian technique (Na\"{i}ve Bayes), machine-learning techniques (support vector machines and back-propagation artificial neural networks) and search-based techniques (genetic programming and artificial immune recognition systems) on FST data collected from two large industrial projects from the telecommunication domain. \emph{Results:} Using area under the receiver operating characteristic (ROC) curve and the location of (PF, PD) pairs in the ROC space, GP showed impressive results in comparison with other techniques for predicting fault-prone modules at both integration and system test levels. The use of faults-slip-through metric in general provided good prediction results at the two test levels. The accuracy of GP is statistically significant in comparison with majority of the techniques for predicting fault-prone modules at integration and system test levels. (ii) Faults-slip-through metric has the potential to be a generally useful predictor of fault-proneness at integration and system test levels.