Receiver operating characteristics (ROC) curve and the area under the curve
(AUC) value are often used to illustrate the diagnostic ability of binary classifiers.
However, both ROC and AUC focus on high accuracy in theory, which may not be
effective for practical applications. In addition, it is difficult to judge which one is better
when the ROC curves are intersect and the AUC values are equal. Decision curve analysis
(DCA) methods improve ROC by incorporating accuracy and consequences. However,
similar to ROC, DCA requires a quantitative indicator to objectively determine which
one is better when DCA curves intersect. A DCA-based statistical indicator named
maximum net benefit (MNB) is constructed for evaluating clinical treatment regimens
rather than just accuracy as in ROC and AUC. As a simple and effective statistical
indicator, the construction process of MNB is given theoretically. Moreover, the MNB
can still provide effective identification when the AUC values are equal, which is proved
by theory. Furthermore, the feasibility and effectiveness of the proposed MNB are verified
by gene selection and classifier performance comparison on actual data.

 Receiver operating characteristics (ROC) curve and the area under the curve
(AUC) value are often used to illustrate the diagnostic ability of binary classifiers.
However, both ROC and AUC focus on high accuracy in theory, which may not be
effective for practical applications. In addition, it is difficult to judge which one is better
when the ROC curves are intersect and the AUC values are equal. Decision curve analysis
(DCA) methods improve ROC by incorporating accuracy and consequences. However,
similar to ROC, DCA requires a quantitative indicator to objectively determine which
one is better when DCA curves intersect. A DCA-based statistical indicator named
maximum net benefit (MNB) is constructed for evaluating clinical treatment regimens
rather than just accuracy as in ROC and AUC. As a simple and effective statistical
indicator, the construction process of MNB is given theoretically. Moreover, the MNB
can still provide effective identification when the AUC values are equal, which is proved
by theory. Furthermore, the feasibility and effectiveness of the proposed MNB are verified
by gene selection and classifier performance comparison on actual data.