Japan Study: AI Predicts Blood Sugar Spikes After Meals

A study from Japan published in the journal Nutrients (2025) shows that artificial intelligence can accurately predict how much a person's blood sugar will rise — or fall — after a meal, using only basic information that most insulin-dependent patients already track every day.

Researchers at the Kyoto Prefectural University of Medicine worked with 58 adults who take multiple daily insulin injections (MDI) — a common treatment approach for both Type 1 and insulin-dependent Type 2 diabetes in the U.S. and abroad. Participants wore continuous glucose monitors (CGMs) and logged their meals and insulin doses. The team then trained a sophisticated AI model — called a transformer, the same underlying technology behind tools like ChatGPT — to predict the peak blood sugar spike and the lowest point (nadir) after each meal.

What the AI Used to Make Predictions

The researchers tested three versions of the model. The most detailed used full nutritional data (carbs, fat, protein, fiber). But here's the key finding: a model using only carbohydrate counts performed just as well. That's significant because counting carbs is already the standard dietary tool taught to insulin users in the U.S.

How Accurate Were the Predictions?

The AI predicted post-meal blood sugar peaks with a margin of error of about 32 mg/dL, and post-meal lows with a margin of about 22 mg/dL. The model's R² accuracy score was 0.58 for both — meaning it explained roughly 58% of blood sugar variability after meals. Results were also evaluated on the Clarke Error Grid, a standard clinical accuracy tool, with most predictions falling in the clinically acceptable Zones A and B.

For context, the American Diabetes Association (ADA) recommends post-meal blood sugar targets of under 180 mg/dL (1–2 hours after eating) for most adults. Predicting whether a patient is likely to exceed that threshold — or swing dangerously low — could help patients and clinicians make smarter, real-time insulin decisions.

Why This Matters for US Patients

Millions of Americans manage diabetes with multiple daily injections, and post-meal blood sugar spikes remain one of the most difficult parts of that management. Current carbohydrate counting methods require patients to estimate doses based on personal experience and imperfect ratios — a process that often leads to unpredictable highs and lows.

This Japanese research suggests that AI tools built into future CGM apps or insulin pump systems could give patients a preview of what their blood sugar is likely to do after a meal — before it happens. Crucially, because the simplified carb-only model matched the full-nutrition model in accuracy, such a tool wouldn't require patients to weigh every gram of fat and protein. Standard carb counting — already widely practiced in the U.S. — could be enough to power it.

While the AI isn't yet available as a consumer product, this research lays important groundwork. As CGM adoption continues to rise in the U.S. — supported by expanded Medicare and insurance coverage — studies like this one point toward a near future where your glucose monitor doesn't just report what your blood sugar is, but predicts what it's about to become.

Original Study: Tominaga, Hamaguchi, Hamaguchi. "Prediction of Postprandial Blood Glucose Variability Using Machine Learning in Frequent Insulin Injection Therapy with a Simplified Carbohydrate Counting Model." Nutrients, 2025. PMID: 41470777. DOI: 10.3390/nu17243832