A GENERATIVE ADVERSARIAL NETWORK-BASED APPROACH TO IMPROVE PRE-EXERCISE HYPOGLYCEMIA PREDICTION IN TYPE 1 DIABETES: OVERCOMING CLASS IMBALANCE IN CLINICAL DECISION SUPPORT SYSTEMS

Exercise-induced hypoglycaemia represents a critical safety risk for individuals with Type 1 Diabetes (T1D), frequently discouraging participation in structured physical activity despite its well-established cardiovascular and metabolic benefits. Although Continuous Glucose Monitoring (CGM) devices provide rich, high-resolution physiological time-series data, clinical datasets for predictive modelling remain severely constrained by small sample sizes and pronounced class imbalance between hypoglycaemic and normoglycaemic events. This thesis proposes and evaluates two Generative Adversarial Network-based data-augmentation frameworks designed to improve the performance of a Bidirectional Long Short-Term Memory (Bi-LSTM) clinical decision support system for pre-exercise hypoglycaemia prediction in T1D patients. In particular, the first proposed framework, mWGAN-GP, implements the Wasserstein GAN with Gradient Penalty incorporating a pairwise diversity loss activated after epoch 1,000 to prevent mode collapse. The second framework, SHOA-WGAN-GP, embeds the Spotted Hyena Optimisation Algorithm directly into the GAN training loop to dynamically adapt the gradient-penalty weight λgp ∈ [5.0, 20.0] and the diversity-loss weight λdiv ∈ [0.0, 0.3] every 100 epochs based on a composite validation fitness function measuring distributional fidelity and synthetic diversity. The work builds upon the DirecNet clinical dataset comprising 47 pre-exercise CGM recordings (34 HYPO, 13 NO-HYPO), which were partitioned using a 70/30 split for training and testing. On the test set, mWGAN-GP achieved an AUC of 90.9%, sensitivity of 72.7%, specificity 75.0%, precision 88.9%, and F1-score 80.0%. SHOA-WGANGP achieved an AUC of 95.5%, sensitivity of 81.8%, a specificity of 75.0%, precision 90.0%, and an F1-score of 85.7%, identifying more hypoglycaemic patients than mWGAN-GP at the same specificity. These results suggest that GAN-based augmentation can improve classifier generalisation beyond physiological augmentation, and that adaptive loss weighting guided by a swarm-intelligence algorithm may provide further benefit. For future developments, both findings could be used in larger cohorts and across repeated training runs.