Predicting Glucose Levels via Disentangled Patterns with Triplet Network

Increasing prevalence of diabetes has driven adoption of wearable technologies for glucose management, with continuous glucose monitoring (CGM) being important for diabetes patients. However, extracting meaningful patterns from CGM data to enhance glucose level predictions remains a challenge. To address this issue, this study proposes a deep learning method of Triplet Network to disentangle CGM pattern based on specific trends and employs a deep learning-based regressor to predict glucose levels. Especially, the disentanglement method was used to enhance the classification process of the specific glucose trends visible in CGM data. This allows CGM regressor to differentiate basic glucose data into meaningful representations, as the disentanglement contributes to the network to focus on clustering salient figures. We used the Glucobench dataset, a collection of real-world CGM data from 5 diverse cohorts, including 5 to 208 participants. Despite the inherent variability in glucose recordings, our model demonstrated robust trend capture and prediction accuracy. We explored multiple window sizes and dense layer configurations, incorporating a 1D convolutional layer to enhance performance. The model achieved its best accuracy with an RMSE of 6.046 mg/dL and outperformed other machine learning models, with an average RMSE of 10.802 mg/dL across all datasets.