The Impact of Clinical Parameters on LSTM-based Blood Glucose Estimate in Type 1 Diabetes

Abstract

Accurate forecasting of blood sugar levels is essential for managing diabetes, especially Type-1 reducing incidences, and diminishing care, costs in patients. In this study, a Long Short-Term Memory Recurrent Neural Network (LSTM) model has been employed to predict blood glucose levels using clinical data. The research focuses on identifying and analyzing several key parameters that play a significant role in determining future blood glucose levels, ensuring a robust and reliable prediction framework. We have considered patient-specific features: Insulin-Sensitivity-Factor (ISF), total daily dose (TDD) of insulin, HbA1C levels, height and weight of a patient, and age and gender while analyzing the prediction performance for Blood Glucose. We thought training LSTM models on a large dataset and studying the most important predictors with their predictive power would be beneficial. The results indicate that including these clinical parameters improves the accuracy of blood glucose prediction and provides valuable information for individuals to control diabetes. This analysis highlights the efficiency of LSTM networks in making use of patient data to improve prediction models, eventually aiding more effective and individualized treatment strategies for Type 1 diabetic patients (T1D). This work also examines the extent to which each parameter influences the prediction of future blood glucose levels, providing deeper insights into their relative impact and significance in the predictive model.