A transformer-graph neural network framework for cyber twin-driven personalized respiratory monitoring

Maintaining healthy breathing is important to health and preventing serious respiratory complications, many of which can result in death. Individuals living with ongoing chronic respiratory disease (asthma, chronic obstructive pulmonary disease, sleep apnea) need ongoing personal patient-centred continuous non-hospital-based monitoring for diagnosis and appropriate treatment. Many traditional existing patient-centred physical health monitoring systems are insufficient due to their limited flexibility, combining multiple sources of sensor data that monitor patients continuously and act in real time to document health states are yet to be invented. This paper proposes a respiratory monitoring system that utilizes a cyber twin (CT) model, and a graph neural network (GNN) function for relational reasoning of the individual based on multi-source data, with a transformer architecture for long-range behaviour modifications based on time series data. The cyber twin continuously updates the physiological state of each individual and predicts what will happen if treatment occurs. This work provides complete results for the comparison of the proposed transformer GNN versus LSTM-CNN being used as evaluation frameworks for established experimental effectiveness, achieving better comparisons for the performance of the model direct interfaces. The proposed approaches provided superior performance over traditional networks with absolute error reductions of 8-13% on respiratory rate (RR) estimates, and F1 score improvements of + 3.1-6.8 points between multimodal respiratory ensembles relative to each other. A graphical figure and table show all the experimental validity and comparisons.