Today's Edition
MARYAM BUKHTAMIN (ABU DHABI)
A team of researchers from the Mohamed Bin Zayed University of Artificial Intelligence (MBZUAI) and Sheikh Shakhbout Medical City has developed two AI systems that hold the potential to help physicians diagnose patients with dementia earlier and more accurately than is presently possible.
Powered by graph neural networks, a new deep-learning architecture, the systems process multi-modal medical data, and produce results that can be interpreted by physicians, a necessity in healthcare where physicians need to understand how AI systems arrive at their results.
The team’s findings were presented at the 28th International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI) in Daejeon, South Korea.
Dementia, which includes Alzheimer’s disease and other disease sub-types such as vascular dementia, affects around 55 million people globally, a number that is expected to triple by 2050. While there is currently no cure, early and accurate diagnosis can slow disease progression and improve quality of life.
“Dementia is one of the most pressing health challenges of our time and these new models can help doctors improve patient care at various stages of the disease,” said Mohammad Yaqub, Associate Professor of Computer Vision and Machine Learning at MBZUAI and co-developer of the systems.
“Through our research, we aim to transform healthcare outcomes and improve the quality of life for patients living with dementia.”
The first model, ClinGRAD, is a graph neural network that classifies patients with dementia into three categories: mild cognitive impairment, vascular dementia, or Alzheimer’s disease. Distinguishing between these sub-types is difficult, and today 20% to 30% of patients are misdiagnosed, leading to early death and reduced quality of life.
ClinGRAD mimics how clinicians interpret data by focusing on known biological indicators of dementia, such as changes in different regions of the brain that can be analysed through magnetic resonance imaging (MRI) and gene co-expression networks. ClinGRAD achieved an accuracy of 98.75% on a benchmark dataset, outperforming existing methods that included a 3D convolutional neural network (84.27%) and a vision transformer (87.12%).
“Our goal is to give clinicians tools that are both accurate and interpretable, so they can trust the results and use them in real-world settings,” said Salma Hassan, a PhD student in machine learning at MBZUAI, and co-developer of the systems. “We embedded clinical reasoning directly into the architecture, building on known biological relationships instead of starting from scratch.”
The second model, MAGNET-AD, uses a graph neural network to detect Alzheimer’s disease before symptoms emerge. It predicts both a patient’s cognitive score (PACC) and the time at which a patient will develop Alzheimer’s disease.
By analysing longitudinal data, including brain scans, genetic data, and electronic health records, it identifies early characteristics of the diasese and has the potential to be used years before clinical symptoms develop.
MAGNET-AD is the first system of its kind to combine spatial, temporal, and multi-modal data in a single interpretable framework. It achieved a concordance index (C-index) of 0.8582, significantly outperforming the next best model (0.8041).
Both ClinGRAD and MAGNET-AD are designed with interpretability in mind, helping to build trust between AI systems and physicians.
“Our work is about empowering doctors with better information, not replacing them,” Hassan said.
Mostafa Salem, postdoctoral research associate in computer vision at MBZUAI and Dr Vijay Ram Kumar Papineni, consultant radiologist at Sheikh Shakbout Medical City, and Dr Ayman Elsayed, consultant radiologist at Sheikh Shakhbout Medical City, contributed to the development of the systems.
In the future, the researchers will test their models on more diverse datasets and expand the models to incorporate additional types of biological data.
