Transformers in health: a systematic review on architectures for longitudinal data analysis

Transformers are state-of-the-art technology to support diverse Natural Language Processing (NLP) tasks, such as language translation and word/sentence predictions. The main advantage of transformers is their ability to obtain high accuracies when processing long sequences since they avoid the vanishing gradient problem and use the attention mechanism to maintain the focus on the information that matters. These features are fostering the use of transformers in other domains beyond NLP. This paper employs a systematic protocol to identify and analyze studies that propose new transformers’ architectures for processing longitudinal health datasets, which are often dense, and specifically focused on physiological, symptoms, functioning, and other daily life data. Our analysis considered 21 of 456 initial papers, collecting evidence to characterize how recent studies modified or extended these architectures to handle longitudinal multifeatured health representations or provide better ways to generate outcomes. Our findings suggest, for example, that the main efforts are focused on methods to integrate multiple vocabularies, encode input data, and represent temporal notions among longitudinal dependencies. We comprehensively discuss these and other findings, addressing major issues that are still open to efficiently deploy transformers architectures for longitudinal multifeatured healthcare data analysis.