Performing Neurological Disease Detection with a Dynamic Graph Based Transformer
As artificial intelligence continues to evolve at an exponential rate, its influence on the
field of healthcare has also grown, with more complex AI models emerging that are capable of disease detection. One of the more recent models that has become prominent for this
application is the Transformer, which is a network architecture that relies on an
attention mechanism, as opposed to traditional recurrent networks, to draw dependencies between inputs and outputs. This research explores the use of a dynamic graph-based Transformer for seizure detection, in which graph features are learned by the model through attention. As opposed to predefined edges and node locations like a standard graph-based Transformer, which restricts the ability to capture nonlinearity, these components of the graph will be learned dynamically through the attention mechanism during training. The use of the attention mechanism for graph construction also allows for result interpretability by constructing attention heat maps, which indicate graph nodes’ influence on each other. Data from the Temple University Hospital seizure corpus was used to train the dynamic graph-based Transformer model across multiple epochs with various hyperparameter combinations, while recording the sensitivity and specificity of each model. The best performing model achieved a sensitivity of 0.3741 and specificity of 0.7784, which was comparable to previous findings on the same dataset using various other model types. The results indicate the viability of a dynamic graph-based approach, while illustrating the ability to improve performance through more thorough preprocessing and training strategies.