Federated learning AI model could lead to healthcare breakthrough
The potential of artificial intelligence (AI) and machine learning (ML) to improve human health cannot be underestimated, but it faces challenges.
One of the big challenges is managing siled data sources, so researchers aren’t able to easily analyze data from multiple locations and initiatives, while maintaining privacy. This is a challenge that can potentially be solved with an approach known as federated learning.
Today in a research report first published in natural medicineAI biotechnology provider OKAY revealed how powerful the federated model can be for healthcare. Owkin, working with researchers from four hospitals in France, was able to build a model with his open-source technology that he believes will have a significant impact on the ability to help treat breast cancer effectively. Owkin’s AI models were able to accurately identify new biomarkers with the potential to improve personalized medical care.
“Owkin is an AI biotech company and we really have this lofty goal, which is to cure cancer,” Jean du Terrail, senior machine learning scientist at Owkin, told VentureBeat. “We are trying to leverage the power of AI and machine learning, in addition to our network of partners, to move towards this goal.”
Owkin is one of hottest biotech startups on the market today. The company raised $80 million in funding in June 2022 from pharmaceutical giant Bristol Myers Squibb, bringing total funding to the unicorn startup, more than $300 million since the company was founded in 2016.
Why Federated Learning is Critical for Advancing AI Healthcare
In medical and clinical studies, there is often a significant amount of personally identifiable information that must be protected and kept private. Researchers as well as hospitals will also often be required to keep certain data within their own organizations, which can lead to information silos and collaboration friction.
Terrail explained that federated learning provides an approach by which ML training can take place across the various patient data information silos located in hospitals and research centers. He emphasized that the approach developed by Owkin does not require the data to actually leave the source facility and that patient privacy is protected.
The federated learning approach is an alternative to using synthetic data, which is also commonly used in healthcare to help protect privacy. Terrail explained that federated learning gives researchers access to real-world data that is secure behind firewalls and is often difficult to access. In contrast, synthetic data is simulated data that may potentially not be fully representative of what can be found in the real world. The risk with synthetic data according to Terrail is that the AI algorithms built with them could potentially not be accurate.
To protect patient privacy, the Owkin approach involves data going through a process known as pseudonymization. Terrail explained that the pseudonymization process essentially removes any personally identifiable information.
The open source software that enables federated learning
Owkin has developed a technology stack for federated learning called Substrate, which is now open source. The Substra project is currently hosted by the Linux Foundation AI and Data Initiative.
Terrail said the Substra platform enables data engineers in hospitals to connect remote sources for ML training. He called Substra a “PyTorch on steroids” application that allows researchers to add functionality to existing machine learning frameworks, such as PyTorch. The additional capabilities enable the federated learning model approach, where data is stored securely and privately in disparate locations.
Substra technology also uses open source Hyperbook immutable ledger blockchain technology. Hyperledger technology allows Substra and Owkin to accurately track all data used. Terrail said Hyperledger is what enables traceability of every operation performed with Substra, which is critical to ensuring the success of clinical efforts. Thanks to traceability, researchers can verify all the steps and the data that have been used. Moreover, it helps to activate Interpretable AI because not all data resides in a black box that no one can audit.
Improving Breast Cancer Treatment Through Federated Learning
Owkin’s teams worked with researchers from four hospitals and were able to train the federated learning model on clinical information and pathology data from 650 patients.
“We trained the model to predict patient response to neoadjuvant chemotherapy, which is the gold standard,” Terrail said. “It’s basically what you give to triple negative breast cancer patients who are in the early stage, but you don’t know if it will work or not.
The research was designed to build an AI that could determine how a patient will react and whether the treatment is likely to work or not. The model could also help refer a patient to other treatments.
According to Thomas Clozel, co-founder and CEO of Owkin, the breakthrough in cancer treatment is based on the success of the federated learning model which is able to collect more data to train AI than has been done before. .
“We want to build federated learning to break down silos of competition and research,” Clozel told VentureBeat. “It’s about human connection and really being able to create this federated network of the best practitioners in the field and researchers who can work together.”
