Realising smarter, more secure healthcare 

MODERN healthcare has become smarter, benefitting from the use of technology like artificial intelligence (AI), in which machine learning (ML) models “learn” how to make decisions based on patterns found in large sets of patient data. This has in turn helped improve the accuracy of medical diagnoses, as well as accelerate the research and development of critical medicines.

However, experts in recent years realised that the traditional process of developing ML applications through the centralised collection of data is insufficient, as effective ML models for healthcare require more data than what would be freely shared due to issues in security and privacy. These challenges have prevented AI from taking the healthcare industry to the next level, where models that achieve clinical-grade accuracy can only be derived from sufficiently large, diverse, and curated datasets.

To democratise AI and reap the benefits of data in healthcare, there is a need for a training method for ML models that is not subject to the risks of sharing sensitive data outside the institution that holds it. Federated learning provides such a method.

Centralised learning no longer sustainable in healthcare 

Centralised learning has long been the traditional norm in AI modelling. This method involves collecting datasets from various locations and devices, then sending it to a centralised location where the ML model training occurs.

This leads to several risks. Firstly, data stored at a single location can be stolen and exposed, creating huge liabilities for the institution responsible for storing it. Secondly, data owners might not even want to share their raw data in the first place. Though the data owners may be willing to have it used for training, the raw data itself may be too sensitive to share.

Security and privacy concerns also make it difficult to scale globally, especially with questions on data ownership, intellectual property (IP), and compliance with regulations pertaining to individuals’ personal data such as Singapore’s Personal Data Protection Act (PDPA). Under the PDPA, the Ministry of Health mandates that access be limited only to doctors and healthcare personnel involved in a patient’s care, as well as the removal of identifying details when data is being used for internal purposes, for example.

The concerns outlined above lead to fewer institutions contributing data. This in turn hinders the ML model from learning from a diverse and augmented set of data obtained from different institutions and geographical locations, which leads to inaccurate and biased data insights.

What federated learning brings to the table

The main idea behind federated learning is to train a ML model on user data without the need to transfer that data to a single location. This involves moving the training computations to the infrastructure at the data-owning institution, instead of moving the data to a single location for training. A central aggregation server is then responsible for aggregating the insights that result from the training computations of multiple data owners.

Federated learning has training iterations performed on local devices, which brings the main benefit of not compromising or exposing the original data when data is in flight. This means that data remains with the owner, while still being utilised to create global insights. Local model parameters resulting from data owner training are sent to a central server, which aggregates them to form the next global model, and later shared to all participants.

In healthcare, federated learning is already making a difference by using state-of-the-art AI to better detect brain tumours. Since 2020, Intel Labs and the Perelman School of Medicine at the University of Pennsylvania (Penn Medicine) are co-developing technology to enable a federation of 29 international healthcare and research institutions led by Penn Medicine to train AI models that identify brain tumours using a privacy-preserving technique called federated learning. Penn Medicine and Intel Labs were also the first to publish a paper on federated learning in the medical imaging domain, particularly demonstrating that the federated learning method could train a model to over 99 per cent of the accuracy of a model trained in the traditional, non-private method.

Building a robust foundation for federated learning starts with trust

With so much riding on data, it is imperative that organisations have a robust data security strategy in place. Key to this is to keep sensitive data in the cloud inside an access-restricted enclave, commonly known as a Trusted Execution Environment (TEE). Privacy protections like these are critical to providing continuous protection of workloads with regulatory requirements or other sensitive data in distributed networks.

As computing moves to span multiple environments – from on-prem to public cloud to edge – organisations need protection controls that help safeguard sensitive IP and workload data wherever the data resides, as well as to ensure that remote workloads are executing with the intended code. This is where confidential computing comes in. Unlike traditional encryption for data at rest or in transit, confidential computing relies on a TEE for enhanced protection and privacy of the code to be executed and the data in use.

Confidential computing means datasets can be processed much more securely, and the risk of attacks can be reduced by isolating code and data from outside incursions. Hardware-based security solutions exist that help protect data in use via application-isolation technology.

With a hardware-based security foundation, previously vulnerable attack surfaces can be strengthened to not only protect against software attacks, but also help eliminate threats against data in use. Organisations can therefore have peace of mind that their ML model can safely use different datasets, and train algorithms with them while remaining compliant with regulations and security.

Future of federated learning 

By enabling ML models to gain knowledge from ample and diverse data that would otherwise be unavailable, federated learning has the potential to bring significant breakthroughs in healthcare, improve diagnosis, and better address health disparities.

While we are still at the beginning of exploring federated learning, it holds great promise by bringing organisations more closely together to collaborate and solve challenging problems, while mitigating issues related to data privacy and security. At the same time, federated learning can stretch its application beyond healthcare, with great possibilities in areas such as Internet of Things, fintech, and more. In fact, Singapore has already begun exploring federated learning : AI Singapore, the national AI research and development programme, has developed an open source platform – Synergos – which aims to make federated learning more accessible.

The future of federated learning is exciting and will bring AI applications to the next level, and even in employing it in healthcare, we are just scratching the surface of its true potential.

The writer is vice president, sales, marketing & communications group and managing director, Asia Pacific territory, at Intel Corporation