A deep learning framework for HbA1c levels assessment using short-term continuous glucose monitoring data

• Published in: Biotechnology and bioprocess engineering Vol. 30; no. 1; pp. 12 - 29
• Main Authors: Han, Bowen, Wang, Yaxin, Li, Hongru, Sun, Xiaoyu, Zhou, Jian, Yu, Xia
• Format: Journal Article
• Language: English
• Published: Seoul The Korean Society for Biotechnology and Bioengineering 01.02.2025; Springer Nature B.V; 한국생물공학회
• Subjects: Biotechnology; Blood; Blood glucose; blood sampling; Chemistry; Chemistry and Materials Science; China; data collection; Deep learning; Diabetes; Diabetes mellitus; Feature extraction; Glucose; Glucose monitoring; glycemic control; glycohemoglobin; Hemoglobin; hospitals; Industrial and Production Engineering; Machine learning; model validation; Monitoring; people; Research Paper; risk; Telemedicine; 생물공학; China; HbA1c levels assessment; Feature importance learning; Continuous glucose monitoring data; Multi-scale feature extraction
• ISSN: 1226-8372; 1976-3816
• DOI: 10.1007/s12257-024-00161-y

Glycated hemoglobin (HbA1c) is a crucial marker for long-term glycemic control, reflecting cumulative blood glucose history over the past two to three months. Elevated levels of HbA1c are significant indicators of increased risk for diabetes-related complications. Typically, HbA1c measurement requires invasive blood collection, which is time-consuming and inconvenient for patients. Recent advancements in continuous glucose monitoring (CGM) technology offer a convenient way to provide continuous and comprehensive blood glucose data. Therefore, a thorough understanding of the correlation between CGM data and HbA1c is essential for accurate interpretation of CGM data and further promoting this technology. Here, we introduce a deep learning solution called HILA, which can extract glucose features from short-term CGM sensor data at different time scales, and combine with manually extracted glucose features to assess patients’ HbA1c levels, thereby reflecting their long-term blood glucose control. Simultaneously, we designed an interpretable feature importance learning mechanism, which assigns weights to manually extracted glucose features, enhances the performance of the model, and reveals the relative importance of different manual features in assessing HbA1c levels. Experiments on a dataset of 1,832 subjects from the Shanghai Jiao Tong University Affiliated Sixth People's Hospital demonstrated that the HILA framework outperforms other machine learning models. This study used two-day CGM data for HbA1c levels assessment, not only enabling the evaluation of long-term glycemic control using short-term blood glucose data but also providing a new perspective for clinicians to effectively interpret short-term CGM data, thereby optimizing diabetes management.