A Self-Attention Deep Neural Network Regressor for real time blood glucose estimation in paediatric population using physiological signals

With the advent of modern digital technology, the physiological signals (such as electrocardiogram) are being acquired from portable wearable devices which are being used for non-invasive chronic disease management (such as Type 1 Diabetes). The diabetes management requires real-time assessment of blood glucose which is cumbersome for paediatric population due to clinical complexity and invasiveness. Therefore, real-time non-invasive blood glucose estimation is now pivotal for effective diabetes management. In this paper, we propose a Self-Attention Deep Neural Network Regressor for real-time non-invasive blood glucose estimation for paediatric population based on automatically extracted beat morphology. The first stage performs Morphological Extractor based on Self-Attention based Long Short-Term Memory driven by Convolutional Neural Network for highlighting local features based on temporal context. The second stage is based on Morphological Regressor driven by multilayer perceptron with dropout and batch normalization to avoid overfitting. We performed feature selection via logit model followed by Spearman's correlation among features to avoid feature redundancy. We trained as tested our model on publicly available MIT/BIH-Physionet databases and physiological signals acquired from a T1D paediatric population. We performed our evaluation via Clarke's Grid error to analyse estimation accuracy on range of blood values under different glycaemic conditions. The results show that our tool outperformed existing regression models with 89% accuracy under clinically acceptable range. The proposed model based on beat morphology significantly outperformed models based on HRV features.