“A single neuron in the brain is an incredibly complex machine that even today we don’t understand. A single ‘neuron’ in a neural network is an incredibly simple mathematical function that captures a minuscule fraction of the complexity of a biological neuron.”

Andrew Ng, deep learning expert


I selected this article for AIMed’s article of the week simply because I do not believe most clinicians will be perusing articles from this journal Sensors. The deep learning gurus of Bengio, LeCun, and Hinton, in the article reviewed last week, mentioned that the future of deep learning will need to address the myriad of shortcomings. The Ahmedt-Ariztizabal article is an interesting followup, therefore, to last week’s deep learning article as it addresses one of these limitations of deep learning.

The authors explain that the structure of physiological recordings are often irregular and unordered, which renders this data difficult to be conceptualized in a matrix such as that seen in convolutional neural networks. The advent of graph neural networks perhaps is at least a partial solution as it is more friendly to the nuances of biomedical and healthcare data. The interacting nodes that are connected by edges of these graph neural networks can factor in “weights determined by temporal associations or anatomical junctions”.

This review paper first outlined the basics of graph neural networks and then described the different types of graph architectures with their applications in healthcare based on: functional connectivity, anatomical structure, and electrical-based analysis. The authors also elucidated why graph-based deep learning is particularly good for medical diagnosis and analysis: the ability to model unstructured and structured relational data such as brain signals and detection/segmentation of organs. Brain graph of fMRI and EEG data, with their complex data structures, are better suited for graph networks as the complex data are not well suited to vector-based representations. The mathematical descriptions of the representative graph neural networks such as ChebNet and graph convolutional networks are daunting but the brief descriptions are useful to read. The section on case studies of graph neural networks for medical diagnosis and analysis is perhaps the most relevant for clinicians to read, especially those who are in the neurosciences and cardiovascular medicine. The authors end this remarkably comprehensive review of graph neural networks in biomedicine with an erudite section on future challenges and directions, specifically graph representation and estimation, graph complexity, dynamicity, interpretability and generalization of graphs.

The full paper can be read here.