Federated Learning is a concept introduced by Google in 2017. In general, training a model, which is the core of ML, requires computing resources that can collect all data in one place and train it. Because ML has the capability to produce more sophisticated results, even scattered data with similar roles are collected and learned over time. However, there are many cases whereby organizations may not want their data sets to be released to the outside world for reasons such as security or asset value.

The same goes for medical data. In this case, rather than collecting the data in one place, the artificial intelligence (AI) model is downloaded and learned in each computing environment that physically holds the data, and the results are collected again to enable large-scale learning without data flowing out.

“If the data of each hospital is well organized, we can obtain a huge amount of clinical data results by collecting and analyzing it in one place. Kakao Healthcare deployed a joint learning platform that can analyze this safely, easily, and systematically, building a large-scale data community involving 16 domestic hospitals participating as of July 2024,” says Hee.

By combining data from multiple subjects under one ML model by using  Google Kubernetes Engine (GKE), Google Cloud was able to create an environment where each hospital could create and operate a large learning federation based on the individual data sets. Based on this foundation, Kakao Healthcare developed a healthcare data platform and Federated Learning environment to process medical data into appropriate data sets and enable joint learning to occur securely and smoothly, and is supplying those data sets to large hospitals in Korea.

However, there were some obstacles, the first being the unity of data. Because each hospital manages data in its own way, the format, storage, and analysis methods are different, and not all data is optimized for storage and analysis.

Although some data such as body temperature, blood pressure measurements, or blood test results should be universally understood, MRI image interpretation, biopsy interpretation, and doctors' medical records are all managed differently at each hospital. Even if the data is standardized, the codes that deal with the disease may differ from hospital to hospital.

Kakao Healthcare and the group of hospitals it collaborates with conducted two joint learning projects. The first was to conduct a test to see whether it was possible to predict recurrence in breast cancer patients. A large-scale analysis was conducted on a total of 25,000 people, including about 15,000 people from four hospitals for the study and about 10,000 people from another hospital for external validation.

“Usually, this scale of data analysis requires extensive data pre-processing and compliance procedures to prepare. With the possibility of labor intensive complications during the actual analysis stage, the task could take more than two years. However, Kakao Healthcare's joint learning only took four months from preparation to the time the results were released. At the same time, it produced much better prediction results since it was not limited to the small amount of data from individual hospitals,” says Hwang.

He explains that Federated Learning has helped Kakao Healthcare achieve surprising results, even from the testing stage. As the amount of data increases, the accuracy and reliability of the data improves, and as a result, more meaningful results are being produced compared to previous studies that included only data within hospitals.

The prediction performance for each participating hospital previously varied from 0.6397 to 0.8362, but the result of Federated Learning exceeded this at 0.8482. By checking the performance based on external validation data that wasn’t used at all for learning, the result was 0.7769, showing a satisfactory performance of 92% of the Federated Learning result.

According to a customer of Kakao Healthcare, Professor Sungeun Song from Korea University Anam Hospital, “Although the development of predictive models using medical data is being carried out at individual hospitals and research charities, there are limitations in that verification and performance evaluation at other institutions are not easy. At the beginning of the research, there were doubts about whether federated learning would be possible to create a collaborative model by effectively learning from medical data accumulated in multiple institutions while protecting the distributed data itself. However, through collaborative research with Kakao Healthcare, we have clearly learned that this is feasible, and we have become convinced that this is the path forward for future medical data research. I believe that more research and great results will be created through joint learning, and that Kakao Healthcare will play a leading role in this process.”

Another customer, Professor Dosang Cho from Ewha Womans University Medical Center says, “Joint learning using the data platform built with Kakao Healthcare surprised us due to its ease of use. Thanks to the vast amount of standardized medical data system that has already been systematically built within the framework of a data alliance created with the trust of multiple organizations, the experience of clutter-free administrative processing, rapid data processing, and easy and intuitive federated learning process was very impressive. It is clear that without a data platform, not only would it have been difficult to experience, but it would also have been difficult to achieve the desired results in a short period of time. We have no doubt that by expanding the number of participating organizations and continuously accumulating high-quality data, we will be able to use it beyond our imagination and produce excellent results.”

Kakao Healthcare and the participating hospitals in this joint learning journey already have plans to expand the scope of research, continue in-depth research with more hospitals, and publish papers on each study based on joint learning. A project to interpret signs for colon cancer is also underway. Although the hospitals are still in the process of forming a union, the data from each hospital has been standardized for analysis, and various studies are being considered.

Kakao Healthcare's data platform is now advantageous not only for joint learning, but also for managing information within hospitals, where digital transformation of treatment systems, ward management, and operations is actively taking place. There are cases where new systems are built to process data for this purpose, and Kakao Healthcare provides about 100 types of data that hospitals can effectively use.