A multimodal deep learning architecture for predicting interstitial glucose for effective type 2 diabetes management

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Endocrinology

Background:

• Accurate blood glucose prediction is vital for diabetes management.
• Continuous glucose monitoring (CGM) provides real-time data but predicting glucose variability remains challenging.
• Existing deep learning models often use unimodal inputs and overlook individual physiological differences.

Purpose of the Study:

• To develop a multimodal deep learning approach for interstitial glucose prediction in type 2 diabetes.
• To integrate CGM data with personalized physiological information from health records.
• To improve the accuracy and generalizability of glucose forecasting models.

Main Methods:

• A multimodal deep learning architecture combining Convolutional Neural Network (CNN) and Bidirectional Long Short-Term Memory (BiLSTM) for CGM time series.
• A separate neural network pipeline processed baseline health records to capture physiological heterogeneity.
• Fusion of CGM and health record data streams for enhanced prediction.
• Prediction of CGM values with 15, 30, and 60-minute horizons using a 5-minute sliding window.

Main Results:

• The multimodal model achieved Mean Absolute Point Error (MAPE) between 6-18 mg/dL for the Abbot sensor and 14-26 mg/dL for the Menarini sensor across prediction horizons.
• Prediction accuracy reached up to 96.7%.
• The multimodal approach demonstrated superior prediction accuracy compared to unimodal methods.

Conclusions:

• The proposed multimodal model effectively predicts interstitial glucose levels by integrating CGM data with individual health records.
• This approach offers a more accurate and generalizable solution for personalized glucose management in type 2 diabetes.
• The findings support the potential of multimodal deep learning for proactive diabetes care.