HGMLA: A Multitask Learning Model for Assessment of HbA1c and GA Levels Using Short-Term CGM Sensor Data

Simultaneously assessing glycated hemoglobin (HbA1c) and glycated albumin (GA) levels provides information on relatively long-term glycemic control over different lengths of time, which is particularly important for diabetes patients with significant glycemic fluctuations. In recent years, continuous glucose monitoring (CGM) technology has provided a convenient approach for comprehensive blood glucose data collection. Despite the widespread use of CGM in clinical practice, we yet have reliable methods to extract meaningful information from short-term CGM sensor data to conveniently assess HbA1c and GA levels in diabetics. In this work, we propose a multitask model called HbA1c and GA multitask levels assessment model (HGMLA) to simultaneously assess HbA1c and GA levels with short-term CGM sensor data to get comprehensive knowledge of relatively long-term glycemic control. Multiscale convolutional kernels and long short-term memory (LSTM) networks are utilized in the model to extract glucose fluctuation features at different time scales. Meanwhile, we transform the 1-D glucose sequence into a 2-D tensor and use 2-D convolution to extract the glucose trend features. In addition, specific CGM manual features are extracted, and multitask learning (MTL) framework is employed to simultaneously assess HbA1c and GA levels. Experiments on a dataset of 1378 subjects demonstrated that the HGMLA is able to achieve better performance on both HbA1c and GA assessment tasks compared with fully connected (FC) neural networks and mainstream machine learning models, such as XGBoost and SVM, with an accuracy (ACC) of 86.59% and 84.42% for two assessment tasks, respectively. Our model enables the assessment of relatively long-term glycemic control in diabetes patients using short-term CGM data, which assists clinicians in formulating more precise treatment options.