Explainable Machine Learning for Real-Time Hypoglycemia and Hyperglycemia Prediction and Personalized Control...

Area of Science:

• Artificial Intelligence in Medicine
• Diabetes Technology
• Predictive Analytics

Background:

• Young adults with type 1 diabetes (T1D) face acute complications from hypoglycemia and hyperglycemia as they manage their own care.
• Continuous glucose monitoring (CGM) provides real-time data for proactive diabetes management.
• Machine learning (ML) can leverage CGM data for early risk prediction and long-term control insights.

Purpose of the Study:

• To develop and evaluate explainable ML models for predicting hypoglycemia and hyperglycemia up to 60 minutes in advance.
• To utilize CGM data and demographic information for risk prediction in individuals with T1D.
• To employ SHAP (SHapley Additive exPlanations) to understand the features driving risk predictions.

Main Methods:

• Trained ML models (XGBoost) using CGM data from 153 T1D participants (over 28,000 days of usage).
• Incorporated short-term, medium-term, and long-term glucose control features, along with demographic data.
• Applied SHAP to identify the most influential features in predicting individual glucose risk.

Main Results:

• XGBoost models demonstrated high performance in predicting hypoglycemia (AUROC: 0.998, Avg. Precision: 0.953) and hyperglycemia (AUROC: 0.989, Avg. Precision: 0.931).
• ML models significantly outperformed baseline heuristic and logistic regression models.
• SHAP analysis identified key contributing factors for individual risk predictions.

Conclusions:

• Explainable ML models offer precise and timely glucose risk predictions, surpassing traditional methods.
• Identifying key risk factors via SHAP aids in understanding individual glucose control and mitigating long-term T1D complications.
• Improving ML model performance is crucial for reducing alarm fatigue and enhancing CGM user experience.