Reactive hypoglycemia probability model

For a long time, athletes and coaches have shared anecdotal reports of experiencing weakness, nausea, and dizziness when eating too close to exercise. Laboratory studies suggested this phenomenon, reactive hypoglycemia, might be related to meal timing, but controlled lab conditions could never capture the complexity of real-world training scenarios.

This is the first study to use real-world continuous glucose monitoring data to confirm what athletes have long suspected: there is an unfavorable timing window for pre-exercise eating.

🧐 However, are all athletes susceptible to reactive hypoglycemia? How likely is it that you will actually experience it in your training life?

Keep on reading for some insights from the largest study ever conducted on this topic.

Courtesy of Supersapiens, shared for the dissemination of scientific findings

Continuous glucose monitoring (CGM) technology uses a small sensor placed on the skin to measure glucose levels in real time, providing readings every minute. Unlike traditional blood glucose meters that require finger pricks, CGM sensors track glucose continuously for up to 14 days, capturing the dynamic changes that occur during exercise, meals, and daily activities.

This technology has radically changed sports nutrition research: large datasets can now be collected from athletes in their natural training environments, rather than in controlled laboratory settings.

Defining hypoglycemia in an athletic context

In this study, reactive hypoglycemia is defined as interstitial glucose, as measured by CGM, dropping below 70 mg/dL (3.9 mmol/L) for at least 5 minutes within the first 30 minutes of exercise.

Some considerations behind this definition:

Performance impact: Athletes may experience symptoms and performance decrements at higher glucose levels than diabetic patients
Exercise context: Physical activity increases glucose utilization, making even moderate drops potentially problematic
CGM accuracy: The 70 mg/dL threshold should account for sensor accuracy while capturing clinically relevant events
Symptom correlation: This level aligns with when healthy individuals typically experience weakness, nausea, and dizziness. However, even though some athletes may experience reactive hypoglycemia, they might not show the typical symptoms.

The 5-minute duration requirement ensures we capture sustained hypoglycemic episodes rather than brief sensor fluctuations.

Real-world example: CGM data showing reactive hypoglycemia

The screenshot below, from the Supersapiens app, shows reactive hypoglycemia in action. You can see:

Post-meal glucose rise: The glucose curve shows an elevation following food intake (🍽️ icon)
Exercise-induced drop: Glucose levels fall sharply after the start of bike training (🚴 icon)
Close to the hypoglycemic threshold: The reading drops to 71 mg/dL, just above the 70 mg/dL threshold, illustrating how close athletes can get to clinically significant hypoglycemia

Supersapiens CGM screenshot showing reactive hypoglycemia: glucose rising after a meal, then dropping during bike training

Real CGM data showing reactive hypoglycemia: glucose rises after eating, then drops during exercise onset

This example shows why timing matters: the athlete ate in what appears to be the unfavorable 30-90 minute window before exercise, and glucose rose after the meal, then dropped rapidly once exercise started.

What this model shows

Using the largest CGM database ever analyzed for this purpose, we developed a predictive model of the relationship between pre-exercise meal timing and reactive hypoglycemia risk.

Our analysis of 6,761 athletes and 48,799 pre-exercise food events showed an unfavorable 30-to-90 minute window before exercise: eating in this window increases the risk of reactive hypoglycemia.

Before this study, advice on pre-exercise meal timing relied on athlete reports and small lab studies. These data add real-world evidence.

Key findings from our data

8.34% average reactive hypoglycemia rate across all events
Peak risk occurs around 60 minutes before exercise
Individual variation: Less than 15% of people experience hypoglycemia in >20% of their events

Interactive probability explorer

The chart below shows the data behind the model. The input is meal timing: how many minutes before the start of exercise the food was eaten. The output is the percentage of exercise events in which reactive hypoglycemia occurred at that timing. Gray dots are the raw percentages for each 1-minute interval; the teal line is a moving average of them. Below the chart, a grid of 100 dots shows the percentage of individuals who experienced hypoglycemia at the selected timing.

How to use: Move the slider to explore different pre-exercise meal timings. Watch how the probability changes, and notice the higher risk in the unfavorable window (highlighted in orange).

Loading hypoglycemia data...

Model construction methods

Data collection and processing

This model is built from a large database of CGM data from healthy, non-diabetic individuals using Abbott Libre Sense Glucose Sport Biosensors in free-living conditions. Key processing steps included:

Data preparation: Linear interpolation for gaps, 3-minute rolling average for smoothing
Hypoglycemia definition: Glucose <70 mg/dL for ≥5 minutes within the first 30 minutes of exercise
Event classification: Minor (<10 min), medium (10-20 min), and major (>20 min) episodes
Timing standardization: 20-minute time bins to ensure uniform distribution across pre-exercise periods

Statistical modeling approach

The probability model uses binomial logistic regression to predict hypoglycemia risk from meal timing:

Linear model: a straight-line relationship between timing and the log-odds of hypoglycemia (the logarithm of the probability that hypoglycemia occurs divided by the probability that it does not)
Non-linear model: a smooth function with 12 degrees of freedom, to capture the curved effect of timing

An analysis of variance (ANOVA) comparing the two models favored the non-linear model, supporting the unfavorable window hypothesis. The non-linear model was also more accurate (62.05% vs 45.1%). Choosing between a simpler and a more flexible model is the same trade-off as in OverfittingM2: the extra flexibility has to earn its place.

Data balancing and validation

Uniform sampling: Equal representation across all time periods
Balanced dataset: Over-sampling minority events to prevent bias
Binary classification: Threshold at p(x) = 0.5 for positive/negative prediction
Performance metrics: Accuracy, precision, recall, and F-score (a single score that combines precision and recall)

Key results and clinical implications

Population-level findings

Analysis of the large CGM database revealed that reactive hypoglycemia occurred in approximately 8% of all logged exercise events preceded by food ingestion. However, individual susceptibility varied widely:

<15% of individuals experienced hypoglycemia in >20% of their exercise sessions
6% of individuals: Susceptible to hypoglycemia regardless of meal timing
8% of individuals: Susceptible to hypoglycemia but timing-dependent (can reduce risk by avoiding unfavorable window)
86% of individuals: Not susceptible to reactive hypoglycemia under free-living conditions

The unfavorable window

The 30-90 minute window before exercise showed peak hypoglycemia risk, with maximum probability when the meal is eaten around 60 minutes before exercise. The model comparison confirmed this non-linear relationship (p < 0.0001).

Glucose falls (≥40 mg/dL drops) occurred in ~52% of events, primarily within the 15-70 minute pre-exercise window, suggesting that symptomatic glucose changes may be more common than threshold-crossing hypoglycemia.

Individual variation and practical applications

The high individual variability aligns with previous smaller studies reporting hypoglycemia rates of 25-94% depending on study conditions. In free-living conditions, the lower overall incidence (8%) likely reflects:

Natural behavior adaptation: Individuals unconsciously avoid problematic timing patterns
Mixed food types: Real-world meals vs. controlled glucose solutions in lab studies
Variable exercise intensities: Self-selected rather than prescribed intensities

Clinical recommendations

For the 8% of timing-sensitive individuals, avoiding the 30-90 minute pre-exercise window can reduce hypoglycemia risk. Meal timing is also the most practical modifiable factor: it is easier to change than other risk factors such as training status or insulin sensitivity.

With a CGM, susceptible athletes can experiment individually with pre-exercise timing and food choices, and find what works best for them.

Biblio

The model and data on this page come from the study by Zignoli et al. 2023, "Association between pre-exercise food ingestion timing and reactive hypoglycemia: Insights from a large database of continuous glucose monitoring data", published in the European Journal of Sport Science (23, 2340–2348). It also shows how large datasets and probabilistic models can reveal timing windows for physiological events during exercise.

For background on the historical confusion surrounding pre-exercise carbohydrate feeding and reactive hypoglycemia, see the review by Jeukendrup & Killer (2011), "The myths surrounding pre-exercise carbohydrate feeding", published in Annals of Nutrition and Metabolism. It discusses the conflicting advice about carbohydrate timing; our CGM-based study adds real-world evidence on timing-specific hypoglycemia risk.

For an accessible explanation of pre-exercise nutrition timing, with clear graphics, see the blog post by sports nutrition researcher Asker Jeukendrup: "What to eat the hour before a race" (MySportScience, 2015). It is not peer reviewed, but it is a useful complement to the papers above.

For a broader view of continuous glucose monitoring in sport and health, see David Lipman's newsletter post "Continuous Glucose Monitors" (Health Performance Nexus). It discusses the use of CGM beyond diabetes management, in athletic performance and metabolic health research.

This work was supported by Supersapiens. The continuous glucose monitoring data and platform access enabled the large-scale analysis of reactive hypoglycemia patterns in free-living conditions.