Leiden University Medical Center (LUMC), Amsterdam UMC, and Centrum Wiskunde & Informatica (CWI) will join forces to develop an innovative line of research in which new forms of explainable Artificial Intelligence (AI) are developed. AI has many possibilities. In the medical domain, it may for instance be used to predict which treatment is most adequate for each patient based on data. However, as not all doctors understand how to interpret such predictions properly, while sometimes this is very important, the current use of this technology at hospitals is still limited. To help AI become more accessible for medical professionals, the research partners received exactly 881,947 euros from the Gieskes-Strijbis Fund.
Black Box
The black of understanding “why is the system concluding this now?” can hinder the widespread use of AI for medical applications, because doctors often want to know why a certain prediction is made. In addition, it is difficult to gain new domain knowledge from a system of which we do not understand what exactly it has learned.
Peter Bosman, senior researcher in CWI’s Life Sciences and Health group: “AI is changing the world, but the possibilities and impact could be even greater if we could better understand and explain to medical experts how a learned model comes to certain conclusions.”
“A lot can be done in healthcare with AI, but its use has not been widely adopted yet”, says Tanja Alderliesten, project leader on behalf of LUMC. Alderliesten and colleagues believe this is partly due to the technology being considered a ‘black box’ for many. “Doctors don’t fully understand how it works. Consequently, they find it hard to blindly trust its predictions”. By developing new AI forms that are more accessible for medical professionals, researchers hope to support its increased use in clinical settings.
This project focuses on creating techniques that cannot just learn explainable models, but can do so by looking at different types of data at the same time. That is a major innovation, because the vast majority of existing techniques are actually only suitable for one type of data. Alderliesten: “AI in healthcare is very suitable for image analysis, such as CT scans. These models can provide insights concerning whether any abnormalities are found on the scan. However, there is much more data available to us that should also be considered, such as DNA profile, gender, weight, medical history. Better predictions can be made by including the whole spectrum of information”.
Evolutionary Algorithms
Peter Bosman leads a long running state-of-the-art research line at CWI focusing on Evolutionary Algorithms, called GOMEA (Gene-pool Optimal Mixing Evolutionary Algorithm). This research offers a new perspective on how AI can be used efficiently for optimization goals. Algorithms from this research line have already been used for the optimization of internal radiation plans for prostate cancer treatments, which led to a new approach for the development of radiation treatment plans that is now used in clinical practice at the Amsterdam UMC. The grant from the Gieskes-Strijbis Fund will make it possible to deepen research on variants of GOMEA that are specifically focused on the learning of explainable AI models, searching for optimal combinations of symbolic expressions, deep learning features, and Bayesian networks to help to establish insightful probabilistic relationships and make better predictions. The funding will particularly enable extensive validation studies based on medical data through collaborations with physicians and other medical experts at LUMC and Amsterdam UMC.
Innovation center for AI
The partners aim to set up a long-running research program and an Innovation center for AI. For now, the researchers’ focus is on oncology patients. In view of the emergence of AI and the rapidly growing interest from the medical world, the time has now come to fully commit to it so that in a few years’ time, medical staff will be ready for the actual translation of these decision support techniques to clinical practice.
More information
