M.S. AAI Capstone Chronicles 2024

glycemic ranges. This model outputs predicted HbA1c values, representing long-term glucose control. Each model was trained using structured procedures to ensure robust learning and generalization.

All datasets were split into training (80%) and validation (20%) subsets. For CNNs and ViTs, the Nutrition 5k dataset was split based on dish IDs to prevent data leakage. LSTM models used time based splitting to preserve temporal order. Models were trained for 10–80 epochs depending on convergence behavior, with early stopping implemented to prevent overfitting. Batch sizes ranged from 16 (for ViTs) to 64 (for LSTMs), balancing computational efficiency and gradient stability. The selection of loss functions was tailored to each model's specific task and requirements. For our LSTM models, we implemented Mean Squared Error (MSE) for regression tasks, as its quadratic nature heavily penalizes larger errors, which is particularly crucial for medical applications where accuracy is paramount. The CNN and ViT models used Mean Absolute Error (MAE) for predicting carbohydrate content, ensuring interpretability of predictions by directly measuring deviations from true values. The MobileNetV3 implementation leveraged transfer learning, with the freezing layer empirically set. Labels were binarized for multilabel classification and selected binary cross entropy as the loss function, to effectively handle multiple class assignments.

10

244