Predicting risk for nocturnal hypoglycemia after physical activity in children with type 1 diabetes

Area of Science:

• Pediatric Endocrinology
• Computational Health Science
• Diabetes Management

Background:

• Nocturnal hypoglycemia is a common concern for children with type 1 diabetes (T1D).
• Daytime physical activity is a significant, yet difficult to predict, risk factor for late post-exercise hypoglycemia.
• Continuous glucose monitoring (CGM) has facilitated machine learning for hypoglycemia prediction.

Purpose of the Study:

• To assess the risk of nocturnal hypoglycemia in children with T1D, specifically considering the impact of prior physical activity.
• To evaluate machine learning models for predicting 9-hour nocturnal hypoglycemia risk in a structured camp setting.

Main Methods:

• Utilized continuous glucose and physiological data from a sports day camp for children with T1D.
• Developed and compared logistic regression, random forest, and deep neural network models.
• Evaluated model performance using the F2 score, prioritizing the detection of false negatives.

Main Results:

• Analyzed data from 13 children over 66 nights, with 18 instances of nocturnal hypoglycemia.
• A random forest model using only glucose data achieved 71.1% sensitivity and 75.8% specificity for predicting nocturnal hypoglycemia.
• The models demonstrated potential in identifying children at risk for nocturnal hypoglycemia based on physical activity.

Conclusions:

• Physical activity significantly influences nocturnal hypoglycemia risk in pediatric T1D.
• Machine learning models, particularly random forest, show promise in predicting nocturnal hypoglycemia risk.
• Predictive models offer valuable clinical support for proactive management and reducing the burden of T1D in children.