This study examines classification methods that can efficiently distinguish between firms flagged during audit report reviews and firms not flagged. This study seeks to overcome, in two respects, the limitations of logistic regression analysis—a classification method widely used in prior research—which assumes only a uniform linear function between the dependent variable and explanatory variables. First, it is necessary to derive the causal relationships existing among the explanatory variables that influence whether a firm is flagged during audit review. By informing decision-makers about which variables affect the audit review flagging through direct or indirect causal relationships with other variables, this approach can support more effective audit review operations. To this end, this study proposes a General Bayesian Network (GBN) and presents the Markov Blanket derived from the GBN. Second, to improve the accuracy of audit review flagging predictions, this study presents an ensemble method that combines the previously used classification methods of GBN, Naive Bayesian Network (NBN), and C5.0. Based on experiments using data on firms flagged and not flagged during audit reviews from 1990 to 1999, it was empirically verified that both methods proposed in this study provide statistically significant results.