Machine Learning-Driven Hypoglycemia Prediction: Extracting Vital Signals from Continuous Glucose Monitoring Data

This project is an open-source simplified reproduction of Fleischer et al.'s 2022 study, focusing on predicting hypoglycemia events from continuous glucose monitoring (CGM) data using time-series feature engineering and ensemble learning models. Developed by AdrianTheweny and hosted on GitHub (link: https://github.com/AdrianTheweny/CGM-hypoglycemia-prediction), it aims to provide early warnings for diabetes patients, turning dense CGM data into actionable insights. Key techniques include trend analysis, rolling statistics, and RUSBoost (a method for imbalanced data) to address the rare nature of hypoglycemia events.

Diabetes affects hundreds of millions globally, with hypoglycemia being a life-threatening acute complication. Traditional finger-prick monitoring is painful and non-continuous, while CGM tech generates high-density time-series data but lacks direct clinical decision support. The challenge lies in identifying early signs of hypoglycemia from these data curves to prevent dangerous events.

The project uses three key feature types:

1. Backtracking window: Past 55-minute to current 12 discrete glucose readings to capture recent trends.
2. Rolling statistics: 60-minute rolling average to smooth noise and reflect average control.
3. Trend feature: 60-minute glucose change slope, a core signal for predicting rapid drops leading to hypoglycemia.

To handle the imbalance (few hypoglycemia events vs normal readings), the project tests multiple models:

• Logistic regression: Baseline with class weight balancing.
• Random forest: 200 trees with adjusted depth and class weights.
• Gradient boosting: Serial weak learners with sample weight for rare events.
• RUSBoost: Random under-sampling + boosting, ideal for rare event detection in medical data.

Evaluation: Uses ROC-AUC (class distinction) and PR-AUC (better for imbalance), plus precision-recall tradeoff (prioritizing recall to avoid missing events) and 5-fold stratified cross-validation. Implementation: Uses scikit-learn for models, imbalanced-learn for imbalance handling, joblib for model persistence. Modular code separates feature engineering, training, and evaluation for reproducibility.

Impact: Early warnings (esp. night-time) can let patients take action before hypoglycemia occurs, potentially saving lives. Future: Need larger diverse datasets, personalized models (insulin sensitivity, diet, exercise), and addressing regulatory, privacy, and ethical issues for clinical deployment.

AdrianTheweny's project bridges academic research and practical implementation, making medical AI accessible. It shows AI's value in turning health data into life-saving insights, emphasizing that tech's ultimate goal is to protect lives and reduce suffering in chronic disease care.