Sponsored By

Boosting Medical Device Safety with Big Data AnalyticsBoosting Medical Device Safety with Big Data Analytics

Use of advanced analytics can lead to fewer recalls and safer medical devices.

August 13, 2015

4 Min Read
Boosting Medical Device Safety with Big Data Analytics

 Use of advanced analytics can lead to fewer recalls and safer medical devices. 

Joy Gandhi

Medical device production is on the upswing. The market has been forecasted to grow tremendously in the next five years. At the same time, devices are trending toward miniaturization and complex functionality. This is pushing the limits to manufacturability and defect detection.

Proof of these challenges in electronics, for example, is the fact that four out of the five devices with the highest number of Class I recalls in the past five years are complex electronic devices. This calls for very powerful means of analyzing complex manufacturing and design data to detect defects and process issues in order to prevent field failures. Big data analytics techniques’ ability to handle today’s tsunami of data and the huge number of manufacturing and design variables in complex devices provides a powerful means to alleviate this problem.

Finding the best manufacturing process conditions to achieve low field defect levels can be addressed by predictive analytics. Machine learning techniques can provide accurate models of field failure rates as function key predictors. The techniques can sift through thousands to tens of thousands of device and manufacturing parameters to identify the variables of importance. Rules discovery techniques, when ran through these key predictors, determine the variable values resulting in low field defect levels. Conventional regression techniques provide less accurate models for a large number of variables, a condition known as the “curse of dimensionality.”

Design of experiments (DOE) is done to identify the best manufacturing process conditions for reduction of defects. The traditional DOE tests a hypothesis based on the known physics of the process. What if we do not know the hypothesis? This is indeed the case for complex devices with thousands to tens of thousands of device and manufacturing variables. A machine learning technique called unsupervised learning can scour through these variables and discover relationships between variables or transforms of variables that are unknown and cannot be discovered any other way.

Most field failures are marginal parts passing specification limits in manufacturing tests and inspections. Qualicent’s worldwide studies show that these units are marginal to many variables interacting in complex ways. Using an anomaly detection technique, the studies demonstrate the detection of these marginal parts from manufacturing data. The algorithms first identify from the thousands of variables those that have the greatest impact on the anomaly. The key variables of importance are then combined in a composite distance that can be used in a manufacturing process control chart to catch product with high likelihood of failing in the field. This prevents the shipment of potential field failures. The technique identifies the top parameters explaining the anomaly crucial in determining the root cause and corrective action for the anomaly.

The above techniques can be applied throughout the product life cycle, beginning at the design and development phases. This approach maximizes the benefits of these techniques by eliminating design-related defects before more costs are incurred in manufacturing the product or, worse, before they become full-blown recalls. Predictive modelling and rules discovery can be performed on a combination of warranty data and manufacturing data from previous designs to determine the best device operating conditions and the best manufacturing process conditions that result in low field failures. The former ensures correct device operation in the field, and the latter ensures manufacturing process control of the key variables that lead to low field defect levels.

The use of historical warranty data is applicable to design updates and revisions on previous products. For new products or technologies for which no historical data exists, unsupervised learning can be used instead. This technique, when applied to a combination of the verification, validation, and pilot production data, can discover key parameters that impact yield or defect levels. The results can provide the “knobs” in manufacturing that engineers can use to increase yield and lower defect levels. This lowers the risks of field failures or recalls during high-volume manufacturing.

Composite distance can be used as part of supplier qualification during development. When applied on supplier production data, it identifies anomalous material that will likely contribute to field failures. The anomalous material can then be removed from medical device manufacturing. The anomaly detection technique improves the effectiveness of first article inspections and supplier acceptance criteria.

The risk of process excursions and anomalies are highest during the pilot manufacturing and high volume ramp. Composite distance is imperative as an in-line process control monitor during pilot production to contain defects that are undetected by conventional screening.

The use cases for advanced analytics discussed here enable companies to leverage the huge potential of their rapidly growing data that had been largely under-utilized over the years. It can lead to billions of dollars in manufacturing ROI, recall cost avoidance, and enhancement of the safety of medical devices.

Don't miss the MEDevice San Diego conference and tech showcase, September 1–2, 2015.

 Joy Gandhi is cofounder and chief marketing officer at Qualicent. Reach her at [email protected].


Sign up for the QMED & MD+DI Daily newsletter.

You May Also Like