Artificial Intelligence for Non-Invasive Glucose Monitoring: ECG-Based Glycaemia Estimation in Type 1 Diabetes

Type 1 diabetes mellitus requires continuous glucose monitoring (CGM) to maintain optimal glycaemic control and prevent adverse events. While CGM devices are effective, they remain minimally invasive, necessitating alternatives for non-invasive blood glucose levels estimation. Recent studies have explored artificial intelligence techniques for blood glucose monitoring using electrocardiogram (ECG) signals, leveraging the cardiac response to glycaemic variations. This study focuses on developing a deep learning model capable of real-time blood glucose level estimation by combining ECG morphological features with heart rate variability (HRV) parameters. The study includes both paediatric and adult patients, expanding upon previous literature that primarily focused on paediatric populations. Results indicate that incorporating HRV features improves predictive accuracy in paediatric patients, whereas, for the adult subject, ECG morphological features alone appear sufficient to achieve high performance. The Clarke Error Grid Analysis reveals that the model achieves clinically acceptable accuracy in more than 96% of cases for the adult patient, and up to 99% for the paediatric patients, although with considerable intersubject variability (ranging from 87% to 99%). These preliminary findings suggest the potential applicability of this technology for continuous and non-invasive glucose monitoring through wearable devices, paving the way for personalised diabetes management.