Hypoglycemia is one of the most critical complications associated with diabetes mellitus, particularly for insulin-treated patients, as it can lead to severe physiological consequences such as loss of consciousness, seizures, cardiovascular complications, and even death if not detected promptly. Although Continuous Glucose Monitoring (CGM) systems are widely used for glucose tracking, they remain invasive, costly, and may not always provide timely or comfortable long-term monitoring. This motivates the need for reliable non-invasive approaches for early hypoglycemia detection.

This thesis investigates the use of electrocardiogram (ECG) signals combined with deep learning techniques for automatic hypoglycemia detection. Since hypoglycemic episodes induce measurable changes in cardiac activity, ECG signals provide a promising alternative for continuous non-invasive monitoring. ECG segments extracted from the D1NAMO dataset were labeled using corresponding glucose measurements and evaluated under multiple experimental

settings, including leave-one-patient-out evaluation, exploratory window-level evaluation, and patient specific modeling. Several imbalance handling strategies were investigated, including class reweighting and synthetic data generation approaches. In parallel, multiple deep learning architectures were evaluated and compared, including recurrent, hybrid temporal, attention enhanced, and residual network-based models. The results demonstrated that residual-based architecture consistently achieved stronger and more stable performance compared with recurrent and hybrid temporal models. The obtained results further showed that strict subject-disjoint patient independent hypoglycemia detection remains highly challenging due to substantial inter patient ECG variability and the limited number of available hypoglycemic events. In contrast, patient specific models achieved substantially stronger and more reliable performance by adapting to individualized ECG characteristics. To further improve interpretability, Gradient-based class Activation Mapping (Grad-CAM) explainability is applied to highlight ECG regions contributing most significantly to model decisions. The resulting activation maps provided qualitative insight into model behavior and demonstrated stable activation patterns across correctly detected hypoglycemic events.

The findings demonstrate the potential of combining residual deep learning architectures with advanced imbalance handling strategies for ECG based hypoglycemia detection. However, additional validation on larger cohorts, external datasets, and prospective real-world studies remains necessary before clinical deployment.