M.S. AAI Capstone Chronicles 2024

Conclusion Across the three datasets analyzed, advanced machine learning methods proved highly effective in addressing critical aspects of diabetes management. The Tidepool dataset supported accurate short-term glucose predictions with LSTM models, while the Nutrition 5k dataset enabled carbohydrate estimation using YOLOv8 and glycemic load predictions via MobileNetV3Large, assisting in meal-time insulin optimization. The DiaTrend dataset provided reliable long-term glucose trend analysis for HbA1c prediction. Although challenges such as missing data, dataset limitations, and variability in extreme cases remain, these models have shown significant potential for real-world application. Future work will emphasize refining the models, expanding the datasets, and integrating these components into a comprehensive diabetes management system to further enhance patient care. Building on our system and the model performance, we’ve identified a few key areas for future improvement. Dataset expansion presents a significant opportunity for improvement, particularly through incorporating more diverse data such as additional cuisine types in Nutrition 5k and extended monitoring periods for both Tidepool and DiaTrend datasets, while addressing class imbalance through synthetic data generation techniques like SMOTE or GANs. Our model refinement strategy would explore advanced architectures, including Informer ProbSparse self-attention mechanism (Zhou et al., 2021) and hybrid transformer-LSTM models for improved temporal predictions, alongside the implementation of ensemble methods to leverage strengths across different model types.

17

251