Every year, new strains of influenza viruses cause hospitalizations, deaths, and economic hardship worldwide. Though vaccines are effective at combating epidemics, the genetic makeup of influenza viruses continuously change, requiring consistent vaccine updates. Since it takes 6-8 months to produce a new vaccine, scientists need the ability to predict which influenza virus to combat in the upcoming year.
However, manufacturing vaccines ahead of the season incurs risk since the prevalent influenza strain is not yet known. Companies face a difficult decision—Do they develop the vaccine sooner, improving their ability to meet the demand but manufacturing a potentially less accurate vaccine? Or do they wait to manufacture the vaccine, increasing vaccine accuracy but risking not being able to meet the demand?
Solution: Vaccine Manufacturing at Risk program
OSTHUS developed the Vaccine Manufacturing at Risk predictive model to reduce the risk that accompanies the proactive manufacturing of flu vaccines ahead of the biannual World Health Organization (WHO) vaccine confirmation.
Our solution provides a range of competencies for data science applications:
Pre-processing unstructured/semi-structured data
Data conformance through reference/master data sources
Statistical analyses and feature engineering of temporal features
Machine learning model development
Machine learning optimization through polyglot analytic notebooks
Cost-based performance metrics for decision support
We implemented two concentric learning cycles in model development—both for the NLP extraction engine component and the overall predictive model. Throughout the learning cycles, OSTHUS worked closely with customer subject matter experts in developing appropriate gold standards, establishing business process-oriented performance metrics, and ensuring appropriate decision support.
Results: Improved prediction accuracy by 6%
OSTHUS’s predictive model successfully identified new influenza strains and their properties. By improving data findability and accessibility, we enabled data-driven decision-making across the organization. As a result, corporate knowledge increased with the ability to leverage data as an asset. Overall, we improved predictive accuracy by 6% and reduced manufacturing risk, subsequently saving millions of dollars per year.
Our solution is one of the first steps toward a future in which influenza surveillance not only monitors viruses but also predicts them. With the ability to predict future influenza strains, the typical tight deadlines of vaccine development can be effectively managed. Additionally, models like this indicate that predictive evolutionary biology can provide solutions to other problems throughout the world.