Borrowing external information to improve Bayesian confidence propagation neural network

A Bayesian confidence propagation neural network (BCPNN) is a signal detection method used by the World Health Organization Uppsala Monitoring Centre to analyze spontaneous reporting system databases. We modify the BCPNN to increase its sensitivity for detecting potential adverse drug reactions (ADRs).

In a BCPNN, the information component (IC) is defined as an index of disproportionality between the observed and expected number of reported drugs and events. Our proposed method adjusts the IC value by borrowing information about events that have occurred in drugs defined as similar to the target drug. We compare the performance of our method with that of a traditional BCPNN through a simulation study.

The false positive rate of the proposed method was lower than that of the traditional BCPNN method and close to the nominal value, 0.025, around the true difference in ICs between the target drug and similar drugs equal to 0. The sensitivity of the proposed method was much higher than that of the traditional BCPNN method in case in which the difference in ICs between the target drug and similar drugs ranges from 0 to 2. When applied to a database managed by Japanese regulatory authority, the proposed method could detect known ADRs earlier than the traditional method.

The proposed method is a novel criterion for early detection of signals if similar drugs have the same tendencies. The proposed BCPNN tends to have higher sensitivity when the true difference is greater than 0.