M.S. AAI Capstone Chronicles 2024

DiaTrend Dataset

The DiaTrend dataset was pivotal in predicting HbA1c levels, with the LSTM model achieving impressive metrics: a Mean Squared Error (MSE) of 0.0623, a Mean Absolute Error (MAE) of 0.1731, and an R² score of 0.9332. Scatter plots comparing predicted and actual HbA1c values indicated close alignment, though a slight increase in variability was observed for higher HbA1c levels (Figure 6). Feature importance analysis using SHAPE highlighted key contributors, such as normalized glucose levels, rolling statistics (mean and standard deviation), and wavelet coefficients, underscoring the model’s ability to capture long-term glucose trends effectively. The model demonstrated strong reliability in HbA1c predictions across a range of values, proving its potential for integration into real-time diabetes management systems. Its strengths included a high degree of generalizability across subjects and strong explanatory power, as reflected by the high R² score. However, some limitations were evident. Predictions for higher HbA1c values showed greater deviations, likely due to underrepresentation in the training data. Additionally, class imbalance, with sparse data for certain HbA1c ranges, affected the model’s ability to predict extreme cases accurately. Future improvements will focus on addressing these challenges by leveraging techniques like SMOTE to handle data imbalance and exploring additional features, such as lifestyle factors, to enhance predictions. Hybrid architectures, including transformers, will also be considered to improve performance, particularly for outlier cases.

16

250