Hypoglycemia risk with physical activity in type 1 diabetes: a data-driven approach

Sahana Prasanna, Souptik Barua, Alejandro F. Siller, Jeremiah J. Johnson, Ashutosh Sabharwal, Daniel J. DesalvoFrontiers Digit. Health 2023 · DOI

TL;DR

This paper tackles the challenge of exercise-induced hypoglycemia in Type 1 Diabetes (T1D) by developing a data-driven approach to predict risk. By analyzing real-world data from continuous glucose monitors and activity trackers, they identified key predictors, offering individuals with T1D a more informed way to manage their activity and minimize dangerous low blood sugar events.

What Problem Does This Paper Address?

Individuals with Type 1 Diabetes (T1D) often face a significant barrier to regular physical activity: the risk of exercise-induced hypoglycemia (low blood sugar). While exercise offers numerous health benefits, the unpredictable nature of blood glucose fluctuations during and after physical activity can lead to dangerous hypoglycemic episodes. This fear can deter individuals from engaging in beneficial activities like walking, running, or cycling. Current recommendations for managing this risk, such as adjusting insulin or consuming carbohydrates, are often based on general guidelines and lack personalized, data-driven insights. The paper aims to bridge this gap by quantitatively understanding the factors that contribute to exercise-induced hypoglycemia and developing a predictive model.

Methods

The researchers employed a linear regression modeling approach to predict hypoglycemia risk. For each of the three common physical activities studied (walking, running, and cycling), they developed a model to predict the average glucose deviation from 70 mg/dl in the two hours following exercise. This deviation serves as a proxy for hypoglycemia risk.

The models incorporated several predictor variables:

  • Glucose value at the start of exercise: The baseline blood glucose level before physical activity.
  • 2-hour pre-exercise glucose adjusted AUC (Area Under the Curve): This novel metric captures the glucose trend in the two hours before exercise, adjusted for the baseline glucose. It's used to quantify "carbohydrate preparation" – how individuals might be proactively managing their glucose levels before activity.
  • Calories Burned: Used as a surrogate for exercise intensity.
  • Activity Duration: The length of the exercise session.
  • Total Bolus Insulin: The amount of insulin taken within two hours prior to exercise.

Datasets

The study utilized data from the Tidepool Big Data Donation Project. This real-world dataset comprises anonymized data from individuals with T1D, including:

  • Continuous Glucose Monitor (CGM) data (measured every 5 minutes).
  • Physical activity data (type, duration, calories burned) from wearable devices.
  • Insulin bolus doses.
  • Carbohydrate intake logs.

The dataset included information from 50 participants with T1D, who had an average of 26.2 weeks of CGM data.

Key Insights and Ideas

  • The "Carbohydrate Preparation" Metric Matters: The 2-hour pre-exercise adjusted AUC emerged as a crucial predictor. This highlights that not just the starting glucose level, but also the trend and proactive management of glucose in the hours leading up to exercise, significantly influences hypoglycemia risk. This suggests that individuals with T1D can actively influence their risk by strategically managing their glucose beforehand.
  • Baseline Glucose is Still King (but not the whole story): The glucose value at the start of exercise was consistently a strong predictor across all activity types, reinforcing the established understanding that starting with adequate glucose levels is vital.
  • Activity Type Nuances: While the core predictors (baseline glucose and adjusted AUC) were significant across walking, running, and cycling, the influence of other factors like duration and intensity varied, suggesting that personalized strategies might be needed for different activities.
  • Real-World Data is Key, but Messy: The study leverages real-world data, which offers valuable insights into actual behavior. However, it also acknowledges the challenges of inconsistent data logging (e.g., carbohydrate intake), necessitating the development of metrics like the adjusted AUC.

Key Results

  • The glucose value at the start of exercise and the 2-hour pre-exercise glucose adjusted AUC were the most significant predictors of hypoglycemia risk across all three studied activities (walking, running, and cycling), with p-values < 0.001.
  • Higher 2-hour pre-exercise adjusted AUC was associated with fewer hypoglycemic events. For instance, the top 25% adjusted AUC group for walking had a mean of 1.09 hypoglycemic events, compared to 2.57 for the bottom 25% (p < 0.0001).
  • The duration and intensity of physical activity, along with 2-hour bolus insulin, showed a weak association with hypoglycemia for walking, running, and cycling.
  • Linear regression models showed moderate predictive power, with R-squared values ranging from 0.399 (running) to 0.539 (walking).

Limitations and Future Work

Limitations: * Limited Demographic Data: The study lacked detailed demographic information (e.g., race/ethnicity, detailed age stratification), which could influence findings. Pediatric data was also limited. * Absence of Heart Rate Data: Heart rate is a known physiological variable that can predict exercise-induced hypoglycemia, but it was not available in this dataset. * Inconsistent Data Logging: Real-world data often suffers from missing or inaccurate logs (e.g., carbohydrate intake), which can affect the precision of certain metrics. * Proxy for Hypoglycemia: The study used average glucose deviation from 70 mg/dl as a measure of hypoglycemia risk, rather than directly measuring "time spent in hypoglycemia," which could be a more clinically relevant outcome. Many exercise events had zero hypoglycemic readings, which would skew results if "time in hypoglycemia" was the sole metric. * Potential for Manual Data Overwriting: The use of real-world data means some exercise data might have been manually adjusted, introducing subjective biases.

Future Work: * Personalized Models: Develop individualized regression models for each participant to improve predictive accuracy. * Incorporate More Predictors: Include demographic information (age, sex), clinical history (duration of diabetes, prior hypoglycemia episodes), and anthropometric measures (BMI) to enhance model performance. * Develop Intervention Technology: Translate these predictive models into a user-friendly app to assist individuals with T1D in managing their exercise and glucose levels safely. * Recompute for Different Populations: Validate findings across diverse populations to assess generalizability. * Utilize "Time in Hypoglycemia": In future studies with more balanced data, use "time spent in hypoglycemia" as the primary clinical outcome. * Automated Data Logging: Future studies could benefit from fully automated logging of exercise parameters for greater accuracy.

Summary updated

BibTeX
@article{prasanna2023hypoglycemia,
  author = {Sahana Prasanna and Souptik Barua and Alejandro F. Siller and Jeremiah J. Johnson and Ashutosh Sabharwal and Daniel J. Desalvo},
  title = {{Hypoglycemia risk with physical activity in type 1 diabetes: a data-driven approach}},
  journal = {Frontiers Digit. Health},
  year = {2023},
  doi = {10.3389/fdgth.2023.1142021},
}

Loading related content...

Figure