Studies of Applications of Expected Information Theory to Overcome Current Problems in Adverse Drug Reaction Analytics

Purpose: Current adverse drug reaction (ADR) signaling methods used in pharmacovigilance have an extensively empirical nature. It is here described how a signal is developed based on the Theory of Expected Information (TEI), in which partially summated zeta functions emerge naturally from integrals of information measures over prior probability densities. This also provides a convenient and better basis for analyzing, comparing, and even combining, other methods, so that this is to some extent also a review and position paper. Methods: Glass Box AI-style methods, akin to high-dimensional data mining and prediction from it, were used. Results: Examples show how hundreds of thousands of reusable knowledge elements obtained from analysis of the FAERs drug reports database contain statistical information that can be inspected for insight and collectively analyzed and used for alerts, as well as for analytical inference and prediction. These elements can be obtained even for single events, and in principle even for joint events that should have been seen, based on prior expectation, but were not seen. Conclusions: The above is shown to represent a more formal approach to ADR Analytics that is smoothly extensible from sparse data. Sparse data appear in (a) early warning signals in a single year’s quarter, and (b) in more detailed analyses involving joined patient demographic, clinical, and other data to assist in stratification, segmentation, and drug interaction warnings.