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

Heike Leutheuser, Marc Bartholet, Alexander Marx, Marc Pfister, Marie‐Anne Burckhardt, Sara Bachmann, Julia E. VogtFrontiers in Medicine 2024 · DOI

TL;DR

This paper tackles the critical challenge of predicting nocturnal hypoglycemia in children with Type 1 Diabetes (T1D), a common and dangerous complication often triggered by daytime physical activity. By leveraging continuous glucose monitoring (CGM) and wearable sensor data, the researchers developed machine learning models to forecast hypoglycemia risk up to 9 hours in advance, offering a valuable tool for proactive management.

What Problem Does This Paper Address?

Nocturnal hypoglycemia, a dangerously low blood sugar level during sleep, is a significant concern for children with Type 1 Diabetes (T1D). Daytime physical activity is a major contributor to this risk, but predicting its delayed impact on overnight glucose levels is complex and challenging. Current prediction models often focus on shorter time horizons or adult populations, leaving a gap in predicting this critical event for children over an extended period. The authors aim to bridge this gap by developing models that can predict the risk of hypoglycemia throughout the entire night, considering the unique challenges and data available from children.

Methods

The study employed a supervised machine learning approach. They experimented with three types of models: Logistic Regression, Random Forest, and Deep Neural Networks (DNNs), including Multilayer Perceptrons (MLPs) and Recurrent Neural Networks (RNNs). The key innovation lies in the comprehensive data integration, combining:

  • Continuous Glucose Monitoring (CGM) data: Providing real-time blood glucose readings.
  • Physiological data from a wearable sensor (Everion): Capturing activity levels, heart rate, and other vital signs.
  • Static participant data: Including age, BMI, and diabetes duration.

The researchers explored various combinations of these data sources as input features. For the baseline models (Logistic Regression and Random Forest), they performed feature engineering, extracting time-series characteristics from glucose data and selecting the top 15 features. For the DNN models, the raw data was used directly. A crucial aspect was the extended prediction horizon of 9 hours, aiming to predict hypoglycemia from bedtime (10 pm) to the early morning (7 am).

Datasets

The study utilized data collected from 13 children with Type 1 Diabetes during a one-week sports day camp. This dataset included:

  • Continuous Glucose Monitoring (CGM) data: Collected from either Freestyle Libre 2 or Dexcom devices.
  • Self-Monitoring of Blood Glucose (SMBG) data: Recorded hourly, during exercise, or when symptoms of hypoglycemia occurred.
  • Physiological data: Captured by a wearable Everion sensor, including 10 selected signals.
  • Participant information: Age, BMI, and diabetes duration.

This dataset, while smaller than some adult-focused datasets, is notable for its focus on children and the inclusion of both CGM and wearable sensor data for a prolonged prediction horizon.

Key Insights and Ideas

  • The extended prediction horizon is crucial: Predicting hypoglycemia 9 hours in advance is clinically more valuable than shorter-term predictions, allowing for timely interventions.
  • Physiological data matters: Incorporating data from wearable sensors significantly improved the performance of the baseline models, highlighting the impact of physical activity on overnight glucose.
  • DNNs show promise for consistency: While not always outperforming baseline models in raw metrics, DNNs exhibited smaller standard deviations in their performance across cross-validation folds, suggesting greater robustness and potential for generalization with larger datasets.
  • Data imbalance is a challenge: The relatively small number of nocturnal hypoglycemia events (18 out of 66 nights) presents a significant challenge for model training and evaluation.

Key Results

The study achieved a best F2 score of 64.4% using a Random Forest model that incorporated time-series characteristics of glucose data. Logistic Regression with glucose, physiological, and static data achieved an F2 score of 63.0%. For the prediction horizon of 9 hours, the Random Forest model using only glucose data achieved a sensitivity of 71.1% and a specificity of 75.8%. While these numbers might seem modest, the authors emphasize their clinical relevance for proactive management. The study also noted that DNNs generally showed smaller standard deviations in performance, indicating more consistent results.

Limitations and Future Work

The authors acknowledge several limitations:

  • Small dataset size: The data from only 13 children and 66 nights limits the generalizability and robustness of the models.
  • Structured setting: The data was collected in a structured sports camp, which may not fully represent real-world daily life.
  • Feature selection for DNNs: While DNNs used raw data, further investigation into feature selection for these models could be beneficial.
  • Adult compendium for MET values: Activity classes were mapped to MET values using an adult compendium, which might not be perfectly accurate for children.

Future work suggested includes:

  • Transfer learning and larger datasets: Combining datasets or using transfer learning to improve DNN performance.
  • Exploring other classifiers: Investigating algorithms like XGBoost and SVM, which have shown good performance in similar studies.
  • Hand-crafted features: Developing more specific physiological features and exploring advanced feature selection techniques.
  • Individual-based models: Evaluating personalized models that consider each child's unique data.
  • Investigating sleep phases: Analyzing the differences in hypoglycemia risk during early vs. late sleep phases.
  • Incorporating insulin and carbohydrate data: Including these factors, which have been shown to improve prediction accuracy.

Summary updated

Abstract

Children with type 1 diabetes (T1D) frequently have nocturnal hypoglycemia, daytime physical activity being the most important risk factor. The risk for late post-exercise hypoglycemia depends on various factors and is difficult to anticipate. The availability of continuous glucose monitoring (CGM) enabled the development of various machine learning approaches for nocturnal hypoglycemia prediction for different prediction horizons. Studies focusing on nocturnal hypoglycemia prediction in children are scarce, and none, to the best knowledge of the authors, investigate the effect of previous physical activity. The primary objective of this work was to assess the risk of hypoglycemia throughout the night (prediction horizon 9 h) associated with physical activity in children with T1D using data from a structured setting. Continuous glucose and physiological data from a sports day camp for children with T1D were input for logistic regression, random forest, and deep neural network models. Results were evaluated using the F2 score, adding more weight to misclassifications as false negatives. Data of 13 children (4 female, mean age 11.3 years) were analyzed. Nocturnal hypoglycemia occurred in 18 of a total included 66 nights. Random forest using only glucose data achieved a sensitivity of 71.1% and a specificity of 75.8% for nocturnal hypoglycemia prediction. Predicting the risk of nocturnal hypoglycemia for the upcoming night at bedtime is clinically highly relevant, as it allows appropriate actions to be taken-to lighten the burden for children with T1D and their families.
BibTeX
@inproceedings{leutheuser2024predicting,
  author = {Heike Leutheuser and Marc Bartholet and Alexander Marx and Marc Pfister and Marie‐Anne Burckhardt and Sara Bachmann and Julia E. Vogt},
  title = {{Predicting risk for nocturnal hypoglycemia after physical activity in children with type 1 diabetes}},
  booktitle = {Frontiers in Medicine},
  year = {2024},
  doi = {10.3389/fmed.2024.1439218},
}
Figure