P-LSTM: A Novel LSTM Architecture for Glucose Level Prediction Problem

Abhijeet Swain, Vaibhav Ganatra, Snehanshu Saha, Archana Mathur, Rekha Phadke

Abstract: We introduce a novel LSTM architecture, parameterized LSTM (p-LSTM) which utilizes parameterized Elliott (p-Elliott) activation at the gates. The advantages of parameterization is evident in better generalization ability of the network to predict blood glucose levels of patients from a real, vetted data set. The parameter of the Elliott activation is learned from the backpropagation steps of the LSTM which reaps the benefits of learning flexible patterns from data using all features and causal features, as the parameter values change in training phase of p-LSTM. The learning of the parameter is also facilitated by fixed point methods on p-Elliott. This leads to better fit and adds explainability in prediction (due to causal features) to the blood glucose fluctuation patterns over time. The coupling of LSTM architecture with p-Elliott leads to superior prediction of glucose levels. It also provides an excellent technique to fit highly nonlinear temporal data, in comparison to the performance of state-of-the-art methods.