M.S. AAI Capstone Chronicles 2024

prediction errors, particularly with respect to outlier glucose readings, which are crucial for managing hypo- and hyperglycemic events (Perlmuter et al., 2008). The reduction in MSE indicates that the LSTM model was more effective at predicting glucose values accurately, which is critical in preventing dangerous health episodes. The LSTM model also achieved an R² score of 0.9930, compared to 0.9224 for the linear regression model, indicating a much better fit to the data and a higher capability for explaining the variance in glucose levels. This enhanced explanatory power highlights the LSTM's strength in capturing the temporal and non-linear dependencies present in the data. The Mean Absolute Error (MAE) was recorded at 0.1254, and the Mean Absolute Percentage Error (MAPE) was 0.0179, indicating that the model's predictions, on average, were within 1.79% of actual values, making it a reliable tool for managing diabetes. The residuals, representing the difference between actual and predicted glucose values, were centered around zero, indicating low bias and validating that the model was not overfitting. The residual distribution in Figure 1.2 showed a roughly symmetric pattern, further supporting the model's robustness and consistency across different glucose levels. This insight is crucial for real-world deployment, where the ability to generalize across various patient profiles is paramount. Figure 2 presents the comparison between actual and predicted glucose values for the LSTM model. The majority of the points lie close to the line of equality, suggesting that the LSTM model delivers strong predictive performance, especially for outlier glucose readings, which is essential for effective management of hypo- and hyperglycemic conditions.

12

246