Predicting HbA1c levels using Bayesian long-short-term memory with attention mechanism for diabetes management

Abstract

Type 2 diabetes is one of the most common and alarming long-term health conditions globally. The prevalence of diabetes is on the rise, with millions of adults reckoned to be living with the ailment due to sedentary lifestyles and traits. Ineffective glycated hemoglobin (HbA1c) measurements and undiagnosed cases have resulted in a high rate of significant complications that characterize the diabetes crisis. HbA1c measurements are often used to assess the levels of blood glucose of patients with type 2 diabetes over a longer period. Thus, it is essential to design a robust model that can accurately predict the HbA1c levels of patients to augment early diagnosis and further treatment. In this study, a Bayesian self-attention bidirectional long short-term memory (Bayesian SA-BLSTM) model is proposed to leverage the time-series data from patients with type 2 diabetes, thereby accurately predicting their HbA1c levels. When compared with traditional LSTM, Light Gradient-Boosting Machine (LightGBM), extreme gradient-boosting machine (XGBoost), BLSTM, and self-attention BLSTM models using the same dataset, our proposed Bayesian SA-BLSTM model exhibited superior prediction power with vigorous credible intervals to reflect the uncertainty in the predicted HbA1c levels. Our finding is a revelation that the Bayesian extension of SA-BLSTM models provides improved predictive performance for effective type 2 diabetes assessment to assist in early detection and subsequent management of the disease. This is an indication that employing Bayesian methods to extend SA-LSTM models can lead to improved predictive performance.