ESO Educational Webinar report: Artificial Intelligence and Stroke Management

The first ESO Educational Webinar of the year, organised on January 28, addressed the role of artificial intelligence in stroke management. Speakers Dr. Giuseppe Reale and Dr. Carlos Molina gave a fascinating introduction to the history, current stage, and potential future of AI in stroke medicine, followed by a lively Q&A moderated by Dr. Ellis van Etten and Dr. Xabier Urra.

First, Dr. Reale familiarised the audience with the essential terminology and definitions related to AI, i.e. “the ability of a machine or a program to perform tasks that require human intelligence, such as learning, reasoning, pattern recognition and problem-solving”. Next, the talk focused on machine and deep learning and their applications in the pathway of a stroke patient. Dr. Reale explained that although AI is currently routinely involved only in imaging analysis, there are several steps in the stroke pathway under active research that could potentially benefit from the use of AI. These include stroke symptom recognition, activation of emergency medical services, radiological assessment beyond automated software use, complication prevention, aetiological evaluation, outcome prognostication, rehabilitation, and follow-up. In addition, the talk addressed the increasing branch of AI-based solutions targeted at patients, digital health, entailing wearables, telemedicine, mobile health apps, and digital therapies.

Next, Dr. Molina continued with the future expectations and uncertainties concerning AI in stroke medicine. He stated the need to first identify areas in which AI could benefit the human-driven clinical work by either optimising treatment selection, reducing workload, minimising human error, or ensuring effective use of data. After setting the goal, the incorporation of AI requires large amount of high-quality data for input, training the system, validation of the algorithm with new data, and application to the clinic.

However, the process is not devoid of obstacles, as highlighted by Dr. Molina. There are still knowledge gaps between clinical and data science, datasets used for AI training are sometimes small and heterogenous, algorithms may lack transparency impairing interpretability of AI-based results, implementation can be challenging, and issues concerning data privacy and ethics are not entirely resolved. Furthermore, as pointed out during the Q&A, there are challenges with global applicability and environmental burden of computational power required by AI. Most importantly, there is still lack of evidence to prove that the use of AI improves clinical outcomes. As one answer to the need of large multicentre datasets for AI training and related data privacy issues, Dr. Molina presented the concept of federated learning architecture in which AI algorithms are transferred between centres, whereas data do not have to move. This requires proper infrastructure and close collaboration which have been initiated in Europe with projects such as Trustroke  and Umbrella. All in all, both speakers emphasised that AI has great potential to assist in stroke management, but human judgment is still essential in complex clinical decisions.

—