SAS® Leads Drug Safety Session at Annual DIA Meeting

By Eric Brinsfield
Earlier this year at the annual Drug Information Association (DIA) conference, I had the pleasure of organizing and leading a session entitled Impact of Data Mining in Pharmacovigilance: Current and Future. I worked with three speakers who I am honored to know. Attendees were clearly interested in the topic and wanted to know more. Personally, this is an exciting area of research for me, because adverse event signal detection and data mining of healthcare and pharmaceutical data fall into the sweet spot for SAS software.

Drug safety and pharmacovigilance actually span many disciplines including medicine, drug development and statistics. Consequently, many pharmacoepidemiologists or safety reviewers, who have medical backgrounds, usually do not have expertise in data mining or predictive analytics. Likewise, programmers and mathematical modelers usually do not have epidemiology backgrounds. With the new public and regulatory focus on drug safety, however, research teams need to add new tools to their toolbox. Sponsors, regulators, and healthcare providers need to look for new ways to detect safety signals earlier and to filter through volumes of information proactively and efficiently.       

As more organizations start evaluating signal detection and data mining, current practitioners show concern that automation and analytics will replace good medical evaluation. This concern was raised during the questions after our session. In reality, data mining and automated signal detection should be combined with existing methods in order to identify new “hunches” or to weed out too many signals. The results of signal detection must be further confirmed by good medical science. 

During my presentation, I attempted to illustrate how a well-designed risk-management plan should encompass all techniques simultaneously and interdependently.  I presented a pharmacovigilance framework that was actually adapted from an article entitled “A Leader’s Framework for Decision Making,” by David J. Snowden and Mary E. Boone in the November 2007 issue of Harvard Business Review. The framework was really intended to facilitate problem solving in a business setting, but due to my recent immersion in pharmacovigilance and signal detection, the adaptation seemed perfect. 

In summary, the framework instructs the decision maker to assign a problem to a specific context. The four contexts are simple, complicated, complex and chaotic. Depending on the context to which a challenge belongs, the problem-solving approach will be different. In my adaption of the framework, I have basically predefined the problems from pharmacovigilance and pre-assigned them to a context.  

The Simple context contains problems that have predefined or predictable responses. They can be delegated to others. They are the world of “known knowns.” We have seen them before and we know what to do with them. 

In the Complicated context, we know there is a problem but we are not sure what it is yet. Basically, this is the context wherein epidemiologists, physicians and statisticians take a signal and try to explain what is causing it. This is the most essential step, which already occurs today. 

The Complex context offers the next level of abstraction in signal detection. Here we apply signal-detection algorithms to detect historical pattern changes or anomalies in adverse-event frequencies. This is where disproportionality scores come into play. Once a signal is detected, it must be routed to the Complicated context. 

Finally, the Chaotic context is the world of speculation and prediction. Using more advanced data mining techniques, such as neural networks and text mining, we attempt to predict a pattern before a strong historical trend is evidenced. New methods and access to healthcare provider data makes this context particularly interesting now. 

The entire framework is summarized in Figure 1, but each context deserves much more in-depth discussion. I presented this to the audience as a proposal. Based on feedback, I hope to refine this and provide more detail over the next few months. I hope you all find it helpful or at least consistent with your experience.



Through its educational offerings and networking opportunities, DIA provides a neutral global forum for the exchange of information critical to the advancement of the drug discovery and lifecycle management processes.”

Within my session, the four presentations progressed from the high-level historical view of drug safety analysis and signal detection to detailed examples of various analytical methods. 

Specifically, the presentations were entitled:

• Speculations on the Problem by John C. M. Wise, MA, Senior Director, Informatics, Daiichi Sankyo.

• A Decision Matrix for Selecting Analytic and Data Mining Methods for Pharmacovigilance by Eric C. Brinsfield, Global Director, Health and Life Sciences, SAS Professional Services.

• Data Mining in Spontaneous versus Active Surveillance Databases: Past, Present, Near Future by Vitali Pool, M.D., Medical Surveillance and Epidemiology, Global Patient Safety, Sanofi-Aventis.

• Making Sense of Pharmacovigilance Data: The Limits of Meta-analysis
Steve Gardner, PhD, Senior Consultant, BioLauncher Ltd.

With 165 attendees still sitting in the room at the end of the session, questions were numerous and interesting. Of the 40 post-session survey respondents, all felt the session should be repeated and expanded next year. One respondent wrote that “I have a better understanding on how to approach data mining.” Overall, the session was well-received, and I hope to lead a similar session next year. I owe my speakers a huge “thank you.”  They did a great job.

Bio: Eric Brinsfield is Global Director, Health and Life Sciences, SAS Professional Services. 

Eric Brinsfield is Global Director, Health and Life Sciences, SAS Professional Services