Dynamic Classification of Imageless Bioelectrical Impedance Tomography Features with Attention-Driven Spatial Transformer Neural Network

Point-of-Care monitoring devices have proven to be pivotal in the timely screening and intervention of critical care patients. The urgent demands for their deployment in the COVID-19 pandemic era has translated into the escalation of rapid, reliable, and low-cost monitoring systems research and development. Electrical Impedance Tomography (EIT) is a highly promising modality in providing deep tissue imaging that aids in patient bedside diagnosis and treatment. Motivated to bring forth an accurate and intelligent EIT screening system, we bypassed the complexity and challenges typically associated with its image reconstruction and feature identification processes by solely focusing on the raw data output to extract the embedded knowledge. We developed a novel machine learning architecture based on an attention-driven spatial transformer neural network to specifically accommodate for the patterns and dependencies within EIT raw data. Through elaborate precision-mapped phantom experiments, we validated the reproduction and recognition of features with systemically controlled changes. We demonstrated over 95% accuracy via state-of-the-art machine learning models, and an enhanced performance using our adapted transformer pipeline with shorter training time and greater computational efficiency. Our approach of using imageless EIT driven by a novel attention-focused feature learning algorithm is highly promising in revolutionizing conventional EIT operations and augmenting its practical usage in medicine and beyond.