The Sydney GI Database — Methodology Explained

The Sydney Glycemic Index Database (Sydney GI Database v3.2, maintained at the University of Sydney) is the single most cited reference for glycemic index values in nutrition research and dietary-app data pipelines, cataloguing GI values for over 4,000 foods and food products tested against internationally standardized human-subject protocols — yet the methodology that generates those numbers involves a level of biological variability that most users of calorie-tracking apps never see. A GI test requires feeding between 8 and 10 healthy human volunteers a precise 50 g available-carbohydrate portion on at least two occasions, measuring venous blood glucose at 15-minute intervals for 2 hours, computing an incremental area under the curve (iAUC), and expressing the food’s iAUC as a percentage of the reference food (glucose or white bread) iAUC from the same subjects. The resulting mean GI value for a single food thus represents 16–20 individual blood-glucose responses — a deliberately small number, given the cost of repeated venous sampling, which produces test–retest coefficients of variation of ±10–15% even in the same laboratory.

The ISO 26642:2010 Protocol: What “Standardized” Means

ISO 26642:2010 is the international standard governing GI measurement, published by the International Organization for Standardization and developed with direct input from the University of Sydney GI Research Service. The standard specifies exact conditions designed to minimize biological and procedural sources of variability.

Subjects must be healthy, non-diabetic adults with fasting plasma glucose below 5.5 mmol/L (99 mg/dL). Individuals with glucose metabolism abnormalities are excluded because their glycaemic responses are systematically different from healthy adults. The test is conducted after a 10–14 hour overnight fast, with no alcohol or vigorous exercise in the preceding 24 hours. Each subject eats the test portion within 15 minutes — a constraint that ensures the glucose absorption rate is not artificially prolonged by slow eating, which would deflate the measured GI.¹

Blood is sampled from a forearm vein at 0, 15, 30, 45, 60, 90, and 120 minutes post-ingestion. The 7-timepoint sampling schedule represents a pragmatic trade-off: more frequent sampling (every 5 minutes) would more precisely characterize the glucose curve shape, but the burden of repeated venipuncture or intravenous cannulation on subjects limits the protocol to intervals that are tolerable for a volunteer population.

The incremental area under the curve (iAUC) is calculated using the trapezoid method, counting only glucose values above the fasting baseline. Values below the pre-meal baseline (hypoglycaemic dips that sometimes occur after high-GI foods due to insulin overshoot) are excluded from the iAUC calculation. This exclusion is methodologically deliberate — it prevents the paradox whereby a food that causes severe hypoglycaemia after an initial spike would appear to have a falsely low GI because the below-baseline portion cancels out the above-baseline area.¹

The reference food — 50 g of glucose dissolved in 250 mL water, or 50 g of available carbohydrate from white bread — must be tested in the same subjects on at least two separate occasions to establish a reliable personal reference iAUC. The GI of the test food is then expressed as: (mean test food iAUC ÷ mean reference food iAUC) × 100.

Available Carbohydrate Portion Size: The 50 g Challenge

The test portion is calibrated to deliver exactly 50 g of available carbohydrate — defined as total carbohydrate minus dietary fiber — not 50 g of food. For foods with low carbohydrate density, this creates test portions that no one eats in practice, which is the most frequently cited limitation of GI as a practical dietary tool.

White bread contains approximately 47 g of available carbohydrate per 100 g, so the test portion is approximately 106 g — about 4 slices. This is a plausible, if large, serving. For watermelon (approximately 6 g available carbohydrate per 100 g), the 50 g available carbohydrate threshold requires approximately 833 g of fruit — well over a kilogram including rind. The watermelon GI of 76 is measured from a portion that exceeds what a single person would consume in a sitting by a factor of three or more.²

This is the foundation of glycaemic load’s clinical utility over GI alone. Glycaemic load (GL) is calculated as: (GI × available carbohydrates in a realistic serving in grams) ÷ 100. For watermelon at a typical 120 g serving: GL = (76 × 7.2) ÷ 100 ≈ 5.5, which falls squarely in the low GL category (below 10). The high GI does not translate to a meaningful glucose burden at a realistic portion size.

The 50 g available carbohydrate convention was chosen because it delivers a glucose stimulus large enough to produce a clear, measurable iAUC curve in healthy subjects while remaining within physiological ranges. Testing with smaller carbohydrate loads (25 g or 30 g) produces flatter curves with larger relative variance, reducing statistical power. The standardization of 50 g available carbohydrate is thus a methodological necessity, not a dietary recommendation.¹

Between-Laboratory Variability and Database Reconciliation

When the same food is tested under ostensibly identical ISO 26642 conditions in Sydney, Toronto, and Boston, the resulting GI values frequently differ by 10–20 units. This inter-laboratory variability is not a failure of the protocol — it reflects real sources of between-population and between-batch variation that the standardization cannot fully eliminate.

Atkinson et al. 2008, published in Diabetes Care as the most comprehensive GI compilation covering 2,480 foods from 25 years of published studies, documented this variability systematically and established reconciliation rules for the database.³ Outliers more than 1.5 interquartile ranges from the median distribution of published values for a food are excluded as probable protocol deviations. When multiple values remain after outlier removal, the database reports the mean with a standard deviation derived from the published values.

The sources of between-laboratory variability include: differences in the specific glucose or bread reference used (affecting the denominator of the GI ratio); differences in the genetic and dietary backgrounds of the volunteer subjects (populations with historically higher starch intakes sometimes show different glycaemic responses to high-starch reference foods); differences in the actual food samples tested across laboratories (a packaged white bread from Australia has different starch gelatinization characteristics than one from the United States, even when both are labeled “white bread”).

The Sydney database handles brand-specific entries from different manufacturing years as distinct records, which is why the database contains multiple entries for the same food category. “White bread, Canadian brand, 1998” and “White bread, Australian commercial, 2006” are separate entries with potentially different GI values reflecting differences in flour protein content, loaf formation, and baking time — all of which affect starch gelatinization and therefore GI.

For apps using GI data, this multi-entry reality means that a single GI integer per food item is always an approximation. The correct implementation stores the mean and standard deviation when multiple measurements are available.

Factors That Shift GI Within the Same Food

The GI of a food is not a fixed molecular property — it is a biological measurement that reflects the rate of starch digestion, which is altered by cooking, processing, storage, and ripeness. The variability within the same food category is often larger than the variability between foods in the same GI band, which has significant implications for how GI data should be applied.

Cooking time and starch gelatinization: Pasta cooked al dente (7–8 minutes) has a measured GI of approximately 38–45. The same pasta fully cooked (14–16 minutes) reaches a GI of 60–65 in the same laboratory.⁴ The mechanism is starch gelatinization: longer cooking disrupts the crystalline starch structure more completely, making the starch chains more accessible to pancreatic amylase, accelerating digestion and producing a faster, higher glucose peak. The GI difference between al dente and overcooked pasta is larger than the difference between al dente pasta and boiled potato.

Ripeness and resistant starch: Green banana GI is approximately 30–35. As the banana ripens, starch converts to sugar and the GI rises to 55–62 at full ripeness and over 65 when spotted.⁴ The resistant starch in unripe banana behaves like dietary fiber, passing through the small intestine largely undigested.

Cooking and cooling (retrogradation): Cooked starch that is cooled undergoes retrogradation — the starch chains re-associate into a more crystalline structure that is more resistant to amylase digestion. Refrigerated cooked rice has a GI approximately 10–15 units lower than freshly cooked rice. Refrigerated cooked potato has a GI of approximately 56 versus 78 when freshly cooked and served hot. This is why cold potato salad has a meaningfully different glycaemic impact than a baked potato, despite identical composition.⁴

These intra-food GI variations are not captured by single database entries. Nutrition apps that store one GI value per food and present it as a precise measurement are obscuring variability that is larger than many between-food differences.

The Reference Food Debate: Glucose vs White Bread

The Sydney database offers GI values referenced to two different standards: pure glucose (GI = 100) and white bread (GI = 100). This dual-reference system generates persistent confusion when values from different publications or databases are compared without checking which reference was used.

The numerical relationship: white bread itself has a glucose-referenced GI of approximately 73 (range 70–78 across studies). Therefore, if a food has a glucose-referenced GI of 55, its white-bread-referenced GI would be approximately 55 ÷ 0.73 = 75. Using the wrong reference scale shifts every GI value in the database by a factor of approximately 1.38 — the ratio of the two reference standards.

Most contemporary nutrition apps use the glucose reference scale following Atkinson et al. 2008, which adopted glucose as the universal reference for the comprehensive compilation.³ This is the appropriate default for any app building on modern GI research. However, many pre-2000 publications used white bread as the reference, and regional databases (particularly some European and Australian sources) have historically used white bread. When a data pipeline ingests legacy data alongside modern data without explicit scale reconciliation, the merged dataset contains systematic errors that cannot be identified without checking each source’s reference food.

The practical implication for app developers: every GI record imported from an external source should have an explicit reference food flag, and scale conversion should be applied before merging records from different source databases.

How Nutrition Apps Should Implement GI Data

The scientifically correct database record for a GI value includes: the mean GI, the standard deviation or coefficient of variation, the number of subjects (n), the reference food scale (glucose or white bread), the year of measurement, the food preparation method, and the country/laboratory of origin. Displaying GI as a single integer — “GI: 54” — implies false precision that the underlying measurement does not support.

Rounding GI values to the nearest 5 units (low: ≤55, medium: 56–69, high: ≥70) better represents the true uncertainty inherent in the measurement. The three-category classification is the one used in clinical settings for this reason — it acknowledges that distinguishing a GI of 54 from a GI of 58 is beyond the resolution of the measurement method.¹

When GI is unavailable — which is true for approximately 60% of foods in any comprehensive app database — imputing from the food category mean is the least-bad fallback. A novel grain product with no published GI can be imputed from the mean GI of similar grain products (approximately 55–65 for most refined grain foods). But the imputed value should be clearly distinguished from a measured one in the UI and in any clinical export — the uncertainty is categorically different.

CalEye’s method page documents which GI values in its database are measured, which are imputed from category means, and which are estimated from macronutrient composition alone. This transparency is not a disclosure formality — it directly affects how the glycaemic load estimates displayed to users should be interpreted, and which meal logs carry the most versus least uncertainty in their GL figures.

References

1. Wolever TMS. “Is glycaemic index (GI) a valid measure of carbohydrate quality?” European Journal of Clinical Nutrition 67 (2013): 522–531.

2. Brand-Miller J, Hayne S, Petocz P, Colagiuri S. “Low-Glycemic Index Diets in the Management of Diabetes: A Meta-analysis of Randomized Controlled Trials.” Diabetes Care 26, no. 8 (2003): 2261–2267.

3. Atkinson FS, Foster-Powell K, Brand-Miller JC. “International Tables of Glycemic Index and Glycemic Load Values: 2008.” Diabetes Care 31, no. 12 (2008): 2281–2283.

4. Brand-Miller J, Wolever TMS, Foster-Powell K, Colagiuri S. The New Glucose Revolution: The Authoritative Guide to the Glycemic Index. 3rd ed. New York: Marlowe & Company, 2007.

5. Jenkins DJ, Wolever TMS, Taylor RH, et al. “Glycemic index of foods: a physiological basis for carbohydrate exchange.” American Journal of Clinical Nutrition 34, no. 3 (1981): 362–366.

6. Venn BJ, Green TJ. “Glycemic index and glycemic load: measurement issues and their effect on diet-disease relationships.” European Journal of Clinical Nutrition 61 (2007): S122–S131.