New way to protect patient data for training AI models

A recent study in South Korea has explored the use of a cryptographic scheme widely used in secure elections to protect patient data used in training AI models. 

Researchers from Asan Medical Center have applied homomorphic encryption (HE) to big data gathered from multiple institutions to test a predictive AI model. HE is an encryption method that allows computations on encrypted data. 

FINDINGS

In a study, whose findings have been published in JMIR Medical Informatics, EMR data of over 300,000 patients from three hospitals – Asan Medical Center, Seoul National University Hospital, and Ewha Women's University Seoul Hospital – were gathered to train an AI model to predict mortality rates within 30 days after surgery. These were encrypted using HE. 

The study proved that HE can be practically applied to combine volumes of data from various sources while preserving privacy and security. 

Findings also suggest that small hospitals can leverage this encryption method to develop their own AI models by safely tapping into data from larger hospitals. These models are likely to have better predictive capabilities than models developed based on raw data from a single institution. 

WHY IT MATTERS

Over a decade since HE was proven in a dissertation in the United States, many heavily regulated sectors such as finance, IT, and healthcare have explored its application to enhance data privacy. Even in elections, this encryption method is increasingly adopted to ensure accurate poll results. 

The research from South Korea also looked into HE, trying to overcome present limitations in healthcare AI development. While having larger, more diverse data is key to improving an AI model's accuracy and applicability, gathering them is a challenge given stringent privacy protection regulations. 

"In this multicenter study, we used cutting-edge HE to protect personal information leakage and data security. Additionally, HE enables operations and predictive modelling on encrypted data, providing an ultimate solution that can completely resolve issues related to personal information leakage and data security. Furthermore, HE provides the 'strongest' security when used appropriately, such as in outsourced computation, wherein HE secures data breaches in computation," the researchers noted in their findings.

THE LARGER TREND

The Korean government has been supporting local medical institutions in pursuing AI research by facilitating collaboration. For example, the Ministry of Health and Welfare launched the Medical Data Utilisation Project which aims to help researchers in digital health connect with hospitals and gain access to their datasets. Last year, it designated five Korean hospitals to become centres of safe medical data utilisation.