Applying Bayesian Networks to Manage Operational Risk
*John F Amrhein, McDougall Scientific Ltd. 

Keywords: Bayes, REMS, RBM, risk management, safety, pharmacovigilance

Two draft guidances for industry, Oversight of Clinical Investigations - A Risk-based Approach to Monitoring (August 2011) and Format and Content of Proposed Risk Evaluation and Mitigation Strategies (REMS) (September 2009) address risk mitigation in operational processes; the former during clinical trial operations and the latter during pharmacovigilance operations. In both operations a suspected or even unknown hazard puts an outcome at risk. For example, in the case of RBM, failure to collect or record eCRF data (missing data) puts the trial's integrity at risk. In the case of REMS, off-label use of a product puts patient safety at risk. Both guidances state or refer to the use of analytics to monitor and evaluate the performance of the operation and the ability to meet documented goals and supporting objectives. The question remains as to HOW to use analytical methods to support these efforts and thereby by facilitate evidence-based decisions.

This talk will propose Bayesian Networks (BNs) as an answer to the question of HOW to leverage data to mitigate operational risk. The power, generality and flexibility of BNs is accepted and used in diverse industries from manufacturing to banking for risk analysis and decision support. Their popularity is due to them being a means of conducting root-cause analysis in a visual and user friendly platform. Root cause analysis presents probabilities for competing causes of adverse events. For example, missing data in a clinical trial may be due to a poorly designed eCRF or a poorly trained site coordinator. Resources will be allocated very differently depending on which is more likely. Likewise, off label use of a product may be due to the actions of a prescribing physician, pharmacist, or patient. Again, resources will be allocated very differently depending on which is the likely cause. A further benefit to BNs is the ability to address possible, but improbable, events that have never occurred. This is accomplished by combining subject matter expertise with collected data in a quantified manner.

This talk will motivate the use of BNs and demonstrate their use and benefits through a small example. The need to have a statistician supporting the implementation of BNs will be made apparent via an introduction to the underlying statistical methods.