A Blood Sugar Model Built in One Country Predicts Another Country Too

Yes, with a real limit. In 327 Midwestern US individuals, a personalized postprandial glucose model trained only on Israeli meals predicted their glucose responses better than carbohydrate or calorie content did. The Israeli-trained model reached R = 0.596, where carbohydrate alone got R = 0.395 and calorie content R = 0.336. When the model was retrained on meals from both groups, performance rose to R = 0.618.

Can a personalized glucose model trained in one population predict responses in another?

That is the question this study was built to answer. It took a model originally developed in an Israeli cohort, froze it, and pointed it at 327 individuals in the Midwestern United States. Each person wore a continuous glucose monitor for six days, logged meals, gave lifestyle and health information, and provided a stool sample for gut microbiome analysis. The model had never seen these people or their food patterns.

The result: prediction was moderate, not perfect, but it cleared the simple food numbers most people rely on. R is the correlation between predicted and measured post-meal glucose. R = 0.596 means about 35% of the variation in meal responses was explained by the Israeli-trained model. Carbohydrate content alone explained about 16%. Calories explained about 11%.

A model built on Israeli meals predicted Midwestern glucose responses better than carb or calorie counts, and it got sharper when it retrained on local people.

This matters because many people still judge a meal by its carbohydrate or calorie label. That label misses most of what actually moves glucose.

Why carbohydrate and calorie counts are weak

Carbohydrate content is a real driver, but it is a single input into a system with many parts. A fixed number of carbs can produce very different glucose curves depending on:

• what else is in the meal: fat, protein, fiber, food structure
• how fast the stomach empties
• how much insulin the pancreas releases
• how sensitive muscle and liver are to that insulin
• what the gut microbiome does with fiber and other residues
• body size, age, baseline glucose and recent activity

A calorie count is even further from the relevant biology. Calories measure energy content, not the rate or shape of glucose absorption. Two meals with the same calories can be a slow lentil dish or a fast rice dish.

This is why the study used a model with many more inputs. The base model included food components and specific features such as the microbiome. It learned patterns from an Israeli dataset, then transferred those patterns to American meals.

What carries over from one population to another

The surprising part is not that the model transferred. It is that it transferred as well as it did despite population differences.

The Israeli cohort and the Midwestern cohort differed in age distribution, diet and microbiome. Those are not trivial differences. Yet the Israeli-trained model still beat carbohydrate and calorie counts in the American data. That suggests the underlying biology is shared: postprandial glucose is driven by similar mechanisms in both groups.

The model's inputs matter here. If it only used carbohydrate, transferring across populations would mean nothing. But because it also includes features like the microbiome, it captures some of the individual variation that travels across diets and populations.

What retraining on local data changed

The R value went from 0.596 to 0.618. That is a small real improvement, not a leap. The paper attributes it to differences in age distribution, diet and microbiome. Retraining lets the model absorb local patterns: how Midwestern meals are composed and which microbes are common.

In practical terms, this means a transferred model is usable, but a locally adapted model is sharper. If this became a product, you would expect the first predictions to be useful, then improve after the model sees enough people like you.

For an individual, that tracks with how continuous glucose monitoring already feels. Your first week with a sensor is partly orientation. After a few weeks, you learn which specific meals spike you, not just which ones should spike you from their carb count.

What the curve looks like

On a CGM trace, a postprandial glucose response is a curve, not a single number. The study computed a postprandial glycemic response from the logged meals, and the reported R refers to that computed response, not to the full glucose curve.

A useful way to picture the idea is that the predicted and measured curves rise and fall together, but not perfectly. If the model predicted a peak at 150 mg/dL, the measured peak might be 135 or 165. The trend is right more often than carbohydrate counting, but individual errors remain.

Do you need a stool test, and what still thins out

No. The model used microbiome data because it improves prediction at the group level. That does not mean you need a stool test to learn from your own glucose.

A CGM gives you the output directly. You eat, log the meal, and watch the curve. That personal response already contains the microbiome, insulin, stomach emptying and all the rest. The main value of a model is filling in predictions for meals you have not eaten yet.

The limits stay in view. Six days of CGM per person is enough to compare models on many logged meals, but it does not show long-term change. R = 0.618 still leaves most of the variation unexplained. Some people will get better individual predictions, some worse. The cohort was Midwestern US individuals, so the findings may not transfer cleanly to children, pregnancy, type 1 diabetes or insulin users, or to very different diets and microbiomes. CGM also lags blood glucose by roughly 5 to 15 minutes and has its own error.

If you want to see your own pattern, start with the free CGM report at healthosx.com/glucose. It reads LibreView, Dexcom Clarity, Stelo, Lingo or SiBionics exports in your browser, no account, and shows your AGP, time in range and biggest rises. No sensor yet? HealthOS Sensor is a month of continuous glucose with the app for $70, no membership, where it ships, at healthosx.com/glucose/sensor.