Federated Learning: A Game-Changer for Secure and… Leave a comment
Modern healthcare has benefitted immensely by harnessing the power of advanced technologies like Artificial Intelligence (AI), in which deep learning (DL) models “learn” to make decisions based on patterns found in large sets of patient data. This in turn has helped improve the accuracy of medical diagnoses, accelerate the research and development of new drugs, and contributed to predictive and preventive medicine.
As models that achieve clinical-grade accuracy can only be derived from significantly large, diverse, and curated datasets, experts in recent years have realized that the traditional process of developing DL applications through the centralized collection of data is a challenging process and limits the extent of datasets. Effective DL models for healthcare require easy access to anonymized patient datasets and also need large datasets for better accuracy. However, concerns due to security and privacy issues create challenges to obtain data from individuals or institutions. Moreover, accessing silos of relevant data spread across different hospitals, geographies, and other health systems, while also facing issues of data storage and complying with regulatory policies, is a massive challenge.
This is where Federated Learning (FL) comes in. The main idea behind federated learning is to move the AI model and required compute to all the locations where the relevant data originates and resides, instead of moving the data to a central location for training. A central aggregation server is then responsible for aggregating the learnings from each of the locations to create the final AI model. Aster DM Healthcare, one of the largest private healthcare service providers operating in GCC and India, collaborated with Intel and CARPL to build a state-of-the-art ‘Secure Federated Data Collaboration Platform.’ The collaboration aims to boost innovation in areas such as drug discovery, diagnosis, genomics, and predictive healthcare, and will also allow clinical trials to access relevant data sets in a secure and distributed manner. To facilitate the adoption of Federated Learning, Intel has led the development of an opensource federated framework called OpenFL. OpenFL along with Intel® Software Guard Extensions (Intel® SGX) provides a secure mode of doing federated learning securing protecting both the data and AI model.
Intel® SGX offers hardware-based memory protection by isolating specific application codes and data in memory. This comprehensive, secure FL solution enables the protection of workload intellectual property (IP) and secures health data with its custodians.
Nivruti Rai, Country Head, Intel India & Vice President, Intel Foundry Services, said, “AI applications are at the cusp of revolutionizing healthcare through timely and effective screening, diagnosis, and treatment of diseases. Getting access to high-quality training datasets and addressing limitations in the form of regulatory frameworks and geographic boundaries are critical imperatives. We are very excited about our collaboration with Aster to address these challenges and deploy a first-of-its-kind Secure Federated Learning Platform in India. It offers a real-world solution by addressing key aspects like security, trust, and confidentiality for optimal use of data. This solution can be potentially offered as a service in collaboration with the healthcare ecosystem to be used by both AI researchers and data custodians in their pursuit of advancing AI innovation – enabling quality and affordable healthcare at scale across the globe.”
The capability of federated learning was demonstrated using hospital data from the Kerala, Bengaluru, and Vijayawada clusters of Aster Hospital. Over 125,000 chest X-Ray images, including 18,573 images selected from over 30,000 unique patients from Bengaluru were used to train a CheXNet AI model. Using Federated Learning to detect abnormalities in the X-Ray report, the 18,573 unique images provided a 3 per cent accuracy boost due to real-world data that was otherwise not available for training the AI model.
Dr Azad Moopen, Founder, Chairman and Managing Director of Aster DM Healthcare, said, “Data is now recognized as the new fuel for quantum leap in all sectors, especially in healthcare. A large part of the real-world data in healthcare now sits in silos. Using technology that transforms data into beneficial statistics we can gain access to larger datasets which will help to develop a personalized healthcare. Aster Innovation and Research Centre which has access to large volumes of data, in collaboration with Intel and CARPL has developed technology to anonymize the patient data and use it. This collaborative platform of India’s first-of-its-kind Secure Federated Learning-based health data platform will unlock opportunities for healthcare ecosystem partners like Pharma and equipment manufacturers. We hope this will exponentially increase the ability to analyze data, develop predictive diagnosis and personalized treatment for patients, while assuring complete data security and confidentiality.”
Globally, Federated Learning has already made a marked difference by harnessing state-of-the-art AI to better detect brain tumors. Since 2020, clinicians from Symbiosis Center for Medical Image Analysis, Pune, National Institute of Mental Health and Neurosciences, Bengaluru, and Tata Memorial Hospital, HBNI, Mumbai, amongst others, have helped Intel and the University of Pennsylvania conduct the medical industry’s largest Federated Learning study. With datasets from 71 institutions across six continents, the study exhibited the ability to improve brain tumor detection by 33 per cent.
As Secured Federated Learning gains momentum, it holds great promise as it allows organizations to synergize and solve challenging problems, while mitigating issues related to data privacy and security. Moreover, its applications aren’t just limited to healthcare, with great possibilities in areas such as the Internet of Things, Fintech, and much more.
