What we did
- Machine learning with Amazon SageMaker
- K-means clustering
- Logistic regression
- Image recognition
When Veripad’s CEO Bishoy Ghobryal goes to visit his family in Egypt, he always brings a large suitcase of medications. Because where his family lives, “sometimes the medication works and sometimes it doesn’t.”
The World Health Organization estimates that one in ten medical products in low- and middle-income countries is either substandard or falsified. That means vital medicines—like antibiotics, pain relievers, and anti-malarial drugs—are often worthless and sometimes deadly.
The big idea
Ghobryal and the other Veripad founders met as biomedical engineering students in college, where they were inspired by an innovation called the Paper Analytical Device (PAD). This chemical test card can identify components of common medications in minutes, with no lab equipment. You just crush one tablet of the medication you want to test, rub it over the card, and soak the card in water.
It solves for very affordable detection of falsified medicine in remote locations. The unique aspect of this technology is its incredible ease of use and portability. You just slip it in your pocket.Bishoy Ghobryal Co-founder and CEO, Veripad
But when a PAD has 12 chemical testing lanes and can detect up to 60 different types of medication, even trained human readers can only “eyeball” it with 80-85% accuracy.
The Veripad founders set out to build a mobile app that anyone could use to photograph a PAD and get clear results. They applied machine learning algorithms to learn from a dataset of PAD images. Their first-generation app achieved 80% accurate classification of medication type, chomping at the heels of human experts.
Make it better
“They chose the right general approach to machine learning, but it wasn’t performing as well as they’d hoped,” says Sandeep Chitta, a Slalom analyst focused on advanced analytics.
Veripad partnered with Slalom on a proof-of-concept to improve the model with Amazon SageMaker. This powerful new platform fast-tracks the process of setting up a machine learning environment and training a solution.
“We developed Amazon SageMaker to remove all of the complexity and inefficiency that’s involved in machine learning to make it easy for developers to easily get started and become skilled building, training, and deploying models, “ says Joel Minnick, AWS’s machine learning and AI marketing leader.
For Veripad’s challenge, we agreed on a modular approach of dividing each PAD image into its 12 lanes and processing them separately. This approach would make the solution more robust and support other PADs in the future. We worked with Veripad to apply image processing techniques like noise cancellation and contrast change to improve the quality of each image. From there, we used clustering techniques to differentiate color patterns and finally logistic regression to classify the type of drug and also assess its purity.
Five weeks later, the new model could “read” a PAD in a matter of seconds and report back with 90% confidence, 5-10% higher than human experts.
“We’re on a good path,” says Jason Ki, Veripad CTO and co-founder. “Our short time with Slalom laid a foundation for us to continue working. We want to get close to 100% accuracy.”
To Kenya and beyond
Veripad is now making technical improvements to its app and gathering data to continue developing the machine learning model. The team is also beginning to mass-produce PAD cards in Kenya, where the counterfeit medication problem is severe.
Soon, they’ll begin several phase-two pilots with small NGOs and National Quality Control Laboratory, a part of the Kenyan Ministry of Health. The team hopes one day Veripad will be used worldwide—not only to detect counterfeit medications, but to gather data that tracks their spread, identifies patterns, and saves even more lives.