Posted by Greg Szwartz, on October 7, 2016
Sometimes, being able to quickly separate critical customer complaints from others is a matter of life and death. Welcome to the daily challenge of the medical device industry.
The volume of complaints can be crushing, especially for a high-profile product. Which are reportable to the FDA? Setting aside the issue of reporting, which represent opportunities for safety and quality improvements? For instance, a patient may complain about something that’s not a safety issue–a broken shipping box. Another client may raise concerns about something far more serious. From an analytics perspective, it can be difficult to distinguish between the two, especially in the face of a large volume of complaints, coming from virtually anywhere in the world, in any language, from any customer or third party.
Difficult, yes–but totally possible. In fact, we’ve done it. We created an application that assembles data from disparate sources into a unified relational database, then uses a mix of statistical and visualization techniques to identify different types of potential safety signals within the data. From there, it’s all served up in a visually oriented user interface that allows users to access, filter and apply rigorous statistical tests on massive volumes of complaints data.
Humans still play a central role in this process. Statistical models tend to overlook the importance of the human element–in this case, people who know the products and who understand how products operate and interact with patients. With the right interface, they can interact with data models in a way that integrates their knowledge naturally and easily. When a “real” signal is detected in the data, an analytics flags it and assigns the signal to an investigator using a workflow tracker. The data set is frozen, the issue is reported to the FDA if necessary, and the company investigates until the case is closed.
Finding problems earlier and faster often means saving lives in this industry–fixing issues before they become larger problems.
For data scientists only: This approach incorporates more than a dozen advanced data mining and statistical techniques, including time series algorithms (such as EWMA, CUSUM, and Shewhart charts), disproportionality analysis (such as PRR and Multi-item Gamma Poisson Shrinker), Poisson regression models, and rare event detection to analyze the entire product lifecycle and myriad product types. For mature products with known defect rates, the platform mainly relies on time series algorithms. For products at the end of their lifecycle, disproportionality analysis is typically the method of choice. For new products, EWMA allows analysis to develop a trend and cycle up to a steady rate fairly quickly, to determine if an identified issue is of broader concern–fast.
If you’d like more content about how our analytics team is helping to hone the craft of data science, subscribe to our content.