Using Swarm Learning approach to aid doctors in treating stroke patients

Researchers at DZNE and the Department of Vascular Neurology at the University Hospital Bonn (UKB) aim to develop a computer model based on artificial intelligence (AI) to aid doctors in treating stroke patients. Serving as a digital assistance system, it is intended to predict the long-term outcome of patients after a minimally invasive treatment (mechanical thrombectomy) and potential complications, thereby helping doctors decide on the best possible therapy. A proof-of-concept study will now be undertaken to determine whether this is feasible using data from the "German Stroke Registry" and additional brain images. The project relies on an AI technology called "Swarm Learning", breaking new ground in the secure analysis of distributed medical data, and aims to lay the foundation for a network of clinics in Germany and beyond. CISPA Helmholtz Center for Information Security is also involved in this endeavor, which is funded by the Helmholtz Association with 250,000 euros.

A stroke is manifested by neurological symptoms, such as speech deficits or paralysis. The most common cause are blood clots: plugs in brain vessels that obstruct blood flow and thus oxygen supply. This situation is referred to as "ischemic" stroke.

In such an event millions of brain cells die every minute unless countermeasures are taken quickly. This is very time-critical. Time is brain, as they say." 

Dr. Omid Shirvani, physician and DZNE scientist

AI for personalized medicine

Possible measures are for example medicinal dissolving of the blood clot or mechanical thrombectomy, a minimally invasive procedure that aims to remove vessel blockage by means of a special catheter. "The type of treatment is decided on a case-by-case basis, depending on factors such as for example the size of the occluded vessel. Based on all available information in an individual case, does thrombectomy have good prospects of success, or does it pose an excessive risk of complications? We aim to develop an AI-based decision-making tool to help with this assessment. It is intended to support doctors who need to act quickly in the event of a stroke. That is our long-term goal. Actual implementation will certainly take some time. But we want to lay the groundwork for this and prove in the current project that our approach does basically work," says Shirvani. He emphasizes: "We don't want a black box, the predictions of our computer model should be comprehensible to doctors, so they can make an informed decision for the benefit of the individual patient. That is, our AI needs to have what is called "explainability" and show the features its assessment is based upon. In addition, clear criteria must be developed to ensure that the AI is applied only to patients whom it can assess with high reliability."

Combining different types of data

AI relies on algorithms being trained on large amounts of data in order to recognize patterns. The larger the pool of training data, usually the better the AI will learn. The researchers therefore intend to combine data from the "German Stroke Registry" with additional brain images generated by magnetic resonance imaging (MRI) or computer tomography (CT). This central registry holds data on the treatment of ischemic strokes from over 20 hospitals across Germany. It contains thousands of cases. "This information comes from the initial examination and follow-up care after a thrombectomy up to three months after intervention. These are primarily detailed entries from the medical records. Associated MRI or CT images of the brain are not included. However, in general, these are kept at the respective hospitals. And there are references in the registry so that images can be clearly assigned", says Prof. Gabor Petzold, Director of the Department of Vascular Neurology at the UKB and Director of Clinical Research at DZNE. "These images contain information that cannot be fully documented in a medical report but which is very valuable for training our AI. That's why we want to link this local data with the information from the central registry."

Traveling algorithm

This is where "Swarm Learning" comes into play. The innovative AI technology is the centerpiece of the current effort. "Traditionally, image data would be collected centrally. However, given the huge amounts of data involved, this is complex and difficult to scale if the network of partners is to grow. And since this is personal data, sharing it requires legal regulations that take a considerable amount of time to comply with. That's why we're taking a different approach. The image data available at the various sites remains local," explains Dr. Anna Aschenbrenner, biomedical scientist at DZNE who is also playing a key role in the project. "This allows us to easily comply with data protection regulations and means that we don't have to move and duplicate large amounts of data. Instead, we send the algorithm to the data via the internet. We let the AI travel from place to place, so to speak, in order to learn. That is the core idea behind Swarm Learning."

Learning collectively

This approach was developed by DZNE in collaboration with IT company Hewlett Packard Enterprise and is currently being applied in various DZNE projects. The term "Swarm" refers to the partners interacting within the network. "With Swarm Learning, everyone involved benefits from the collective data pool without having to share their own data. This data remains on site and confidential in accordance with data protection regulations. This is because the algorithm only extracts parameters without any personal references," explains Prof. Joachim Schultze, Director of Systems Medicine at DZNE, who is also a professor at the University of Bonn. "The result is a trained AI that has learned at all network nodes. It has assimilated the collective knowledge and can even evolve as new data is introduced. In our specific case, we would then have an AI-based computer model that could support doctors in treating strokes. All network partners could use this tool. Regardless of whether they have large or small amounts of their own data, they would all benefit equally from participating in the swarm."

International perspective

Starting with three clinics, including the University Hospital Bonn, the researchers intend to gradually expand their approach to other members of the "German Stroke Registry". For testing purposes, they will start with multicentric data from the "German Stroke Registry" available in Bonn and use it to simulate a swarm in DZNE's computing center before transferring the system to geographically separate locations. "We want to lay the foundation for a nationwide network," says Aschenbrenner. "Furthermore, we are already in talks with partners in the UK to continue our concept internationally. I think there is a lot of room for development."