Meal Composition and Post-Meal Glucose Response: GIVEMEA Study Guide
GIVEMEA Study Guide · Digital Biomarkers (Module 3)
Meal Composition and Post-Meal Glucose Response
Individual assignment 375-0003-00L · CGMacros open dataset (Das et al., 2025) · Review of methods, procedures and results
Secondary analysis of open CGM data45 adults · about 10 days eachCorrelations · mixed models · predictionR and Python
45Adults
687,580Minute rows
1,706Logged meals
1,173Meals analysed
+6.0mg/dL·h per +10 g carbs
0.11Out-of-sample R² (macros)
Central finding
Carbohydrate raises and fiber lowers the glucose rise after a meal, but meal composition alone predicts a new person’s response poorly (out-of-sample R² 0.11); knowing the person helps far more.
The three assignment questions
(1) Basic descriptives. (2) The features extracted from the CGM and meal logs, and why. (3) The correlations found. The report answers them with the open CGMacros dataset instead of self-collected data.
What a “meal response” is
For each logged meal, take the mean glucose in the 15 minutes before as the baseline, then measure the area above it, the peak rise and the time to peak over the next 2 hours. Everything else in the report relates meal features, glycemic group and person to that response.
How to use this guide
Flip through Key Terms first, read Methods and Results out loud as if explaining them to someone, then take the section quizzes. Quiz questions are written to test the meaning of a number, not just recall of it.
What it can and cannot claim
It supports analytical validation only: the measure behaves sensibly and separates groups. It does not show clinical validation or utility, so claims about what a specific food does to a specific person are premature.
Table 2 from the report: the three groups
Normoglycemic
Prediabetes
T2D
Participants
15
16
14
Age (years)
38.7
51.6
54.2
HbA1c (%)
5.2
6.0
7.2
Mean glucose (mg/dL)
122
135
166
Glucose SD (mg/dL)
20.8
26.7
40.4
Time 70 to 140 mg/dL (%)
79.6
66.4
33.2
Time 70 to 180 mg/dL (%)
96.5
90.5
68.1
Meals analysed
373
437
363
Median 2-h iAUC (mg/dL × h)
29.2
42.1
60.3
Mean peak rise (mg/dL)
43.5
54.9
71.8
Tap a card to flip it. Try to say the definition out loud before you flip. ■ Data and measurement■ Meal-response features■ Statistics■ Digital biomarker concepts
Data and measurement
CGM (continuous glucose monitor)
tap to define
CGM (continuous glucose monitor)
A wearable sensor that reads glucose in the fluid under the skin around the clock. The CGMacros files hold one row per minute.
Dexcom vs Libre
tap to define
Dexcom vs Libre
Two sensors worn together. Dexcom (abdomen) was more complete and is used for every glucose feature; Libre (upper arm) is only a check and reads about 32 mg/dL lower on average.
HbA1c
tap to define
HbA1c
A laboratory marker of average glucose over recent months. Here it splits the 45 people into three glycemic groups.
Glycemic groups
tap to define
Glycemic groups
Normoglycemic: HbA1c below 5.7%. Prediabetes: 5.7 to 6.4%. Type 2 diabetes (T2D): 6.5% or higher. Sizes 15, 16 and 14.
Time in range
tap to define
Time in range
The share of readings inside a band. This report uses 70 to 140 mg/dL (tight, near-normal) and 70 to 180 mg/dL.
Free-living data
tap to define
Free-living data
Recorded in everyday life, not in a lab. Realistic, but messier: missing minutes, estimated meals, other activity.
CGMacros
tap to define
CGMacros
An open PhysioNet dataset: 45 adults, about 10 days each, CGM glucose, logged meals with estimated macronutrients, and laboratory HbA1c.
Logged meal
tap to define
Logged meal
Any row where a meal type is recorded. Its calories, carbs, protein, fat and fiber are estimates, so they contain error.
Meal-response features
Pre-meal baseline
tap to define
Pre-meal baseline
Mean Dexcom glucose in the 15 minutes before eating. It must be controlled, because the starting level limits how far glucose can rise.
iAUC (incremental area under the curve)
tap to define
iAUC (incremental area under the curve)
The area above the baseline in the 2 hours after a meal, in mg/dL × hours. One number for the size and duration of the rise.
Peak rise
tap to define
Peak rise
Highest glucose in the 2 hours minus the baseline, in mg/dL. Captures height; correlates 0.95 with iAUC here.
Time to peak
tap to define
Time to peak
Minutes from the meal to the highest value. Only weakly related to any meal feature in this report.
Isolated meal
tap to define
Isolated meal
No other logged meal in the 90 minutes before or the 120 minutes after. Lets a response be attributed to one meal.
Coverage rule
tap to define
Coverage rule
A meal is kept only if at least 80% of the 2-hour window after it has glucose readings.
Coefficient of variation (CV)
tap to define
Coefficient of variation (CV)
Glucose SD divided by the mean. Describes how big the swings are relative to the average level.
Fiber per gram of carbohydrate
tap to define
Fiber per gram of carbohydrate
A derived ratio. Fiber slows absorption, so its share of the carbohydrate matters. It had the strongest correlation with iAUC (−0.31).
Statistics
Spearman ρ
tap to define
Spearman ρ
A rank correlation: do higher values of one variable go with higher values of the other? It copes with skewed data and outliers.
Within-person correlation
tap to define
Within-person correlation
Each value minus that person's own average, then correlated. Removes differences between people so only meal-to-meal variation within a person is left.
Mixed-effects model
tap to define
Mixed-effects model
A regression with fixed effects for what you care about (carbs, fiber, group) and a random effect for each person, because one person's meals are not independent.
Random intercept
tap to define
Random intercept
Each participant gets their own starting level, so a person who responds strongly overall does not distort the estimates for everyone else.
Fixed effects
tap to define
Fixed effects
The average effects shared by everyone: for example, the change in iAUC for each extra 10 g of carbohydrate.
ICC (intraclass correlation)
tap to define
ICC (intraclass correlation)
The share of variation that sits between people rather than between meals of the same person: 0.19 before adjustment, 0.35 after.
95% confidence interval
tap to define
95% confidence interval
A range of plausible values for an estimate. If it includes 0 (for example fat: −3.5 to 0.4), the effect is not clearly different from none.
Marginal vs conditional R²
tap to define
Marginal vs conditional R²
Marginal: variance explained by fixed effects alone (28%). Conditional: fixed effects plus each person's own average level (53%).
Out-of-sample R²
tap to define
Out-of-sample R²
How much of the variation in held-out data a model explains. 0 means no better than predicting the average; negative means worse.
Leave-people-out cross-validation
tap to define
Leave-people-out cross-validation
Train on some people, test on people the model never saw, and rotate until everyone has been held out once (9 groups here).
Bootstrap by participant
tap to define
Bootstrap by participant
Resample whole people with replacement, 500 times, and recompute R² each time to get a 95% interval. The person, not the meal, is the unit.
Random forest
tap to define
Random forest
A flexible prediction method that averages many decision trees. It can capture non-linear patterns but is harder to interpret.
Population vs personalised model
tap to define
Population vs personalised model
Population: fitted on other people. Personalised: fitted on that person's own other days.
Leave-one-day-out
tap to define
Leave-one-day-out
Hold out every meal from one day of a person, train on their other days, predict the held-out day, repeat.
Multiple comparisons
tap to define
Multiple comparisons
The more tests you run, the more likely some look significant by chance. The report did not correct for this.
Digital biomarker concepts
Digital biomarker
tap to define
Digital biomarker
A health-related measure derived from digital device data, here the post-meal glucose response computed from CGM and meal logs.
Analytical validation
tap to define
Analytical validation
Does the measure behave sensibly and repeat? This is the only level of validation the report claims.
Clinical validation
tap to define
Clinical validation
Does the measure predict a health outcome? Not tested in this report.
Utility
tap to define
Utility
Does acting on the measure improve anything for patients? Not tested in this report.
Measurement error (attenuation)
tap to define
Measurement error (attenuation)
Noise in estimated macronutrients pulls the estimated effect of carbohydrate toward zero.
Standardised test meal
tap to define
Standardised test meal
The same meal given to everyone. The identical 712 kcal breakfast eaten 88 times separated the three groups, so it is a candidate response biomarker.
Unmeasured context
tap to define
Unmeasured context
Activity, sleep, stress, the previous meal and medication were not modelled but can all change a glucose response.
Hybrid model
tap to define
Hybrid model
The population prediction plus the person's own average residual. It reached 32% in the within-person test.
Study design at a glance
A secondary analysis of an open dataset (CGMacros). Each logged meal is linked to the 2 hours of glucose after it; meal features are related to that response with correlations and mixed-effects models, and then tested for how well they predict the response of people the model has not seen.
Data source and sample
CGMacros (Das et al., 2025, PhysioNet): 45 adults, about 10 days of free-living Dexcom G6 Pro glucose, logged meals with estimated macronutrients, laboratory HbA1c.
29 women and 16 men; mean age 48 (SD 13, range 18 to 69); mean BMI 31.1; mean HbA1c 6.1%.
687,580 one-minute rows; median 10.6 days per person (range 7.2 to 19.2); 1,706 logged meals (median 38 per person, range 17 to 64).
The files come in eight column layouts, so some variables exist for only some people (METs for 34, a 0 to 3 intensity score for the other 11).
From a logged meal to a response (features)
Baseline = mean Dexcom glucose in the 15 minutes before the meal.
iAUC = area above the baseline over the next 120 minutes (mg/dL × hours). Only the part above baseline counts.
Peak rise = highest value minus baseline; time to peak = minutes to that value.
Derived meal variables: carbohydrate share of energy (4 × carbs ÷ total energy) and fiber ÷ carbohydrate. Also meal type and hour of day.
Per-person CGM metrics: mean glucose, SD, CV, time at 70 to 140 and 70 to 180 mg/dL.
Cleaning the meal log
1,706 logged meals.
Minus 174 with implausible nutrition: under 50 kcal, a macronutrient above 176 g, fiber above carbohydrate, or calories more than 40% below or 60% above what the macros imply.
Minus 37 with less than 80% glucose coverage in the 2 hours after eating.
Minus 322 that were not isolated (another meal within 90 min before or 120 min after).
= 1,173 meals. All 45 people remain (each keeps 15 to 35 meals); only meals were removed.
Correlation and model steps
Spearman correlations between meal features and iAUC, pooled over 1,173 meals, plus a within-person version.
Mixed model with a random intercept per person. Carbs, protein, fat and fiber are entered per 10 g; adjusted for pre-meal glucose, glycemic group and meal type (lunch = reference).
Empty model first, to get the ICC (how much variation lies between people).
Sensitivity checks: keep overlapping meals (1,495 meals), use the Libre sensor, use a 3-hour window.
Prediction steps
Population models: carbs only; four macros; macros + pre-meal glucose + hour; adding HbA1c, BMI and age; and a random forest on the same inputs.
Each person held out once in 9 groups; out-of-sample R² with a 95% interval from resampling participants 500 times.
Personalised test: leave-one-day-out within each person (carbohydrate and fiber only), compared with the population model, the person's average, and a hybrid.
For each result: say what the number is, what it means in plain words, and what it does NOT mean. Then try the quiz.
Carbohydrate raises the response
Each extra 10 g of carbohydrate raised the 2-hour iAUC by 6.0 mg/dL × h (95% CI 5.0 to 7.0), adjusted for pre-meal glucose, glycemic group and meal type. Against a between-meal SD of 55, one 10 g step is a modest nudge, but meals vary by tens of grams.
Fiber lowers it, but handle with care
Each extra 10 g of fiber lowered iAUC by 15.5 (CI −21.7 to −9.3). Because fiber is partly inside the carbohydrate figure, this is the extra effect of fiber at the same carbohydrate. It rests on a minority of meals with more than 10 g of fiber, so do not quote it as a dose.
Protein small, fat unclear
Protein: −2.8 per 10 g (CI −3.9 to −1.7). Fat: −1.6 (CI −3.5 to 0.4). The fat interval includes 0, so there is no clear effect.
Meal type and starting glucose
Breakfasts produced 24.0 mg/dL × h more than lunches (CI 16.6 to 31.4); dinners and snacks did not differ from lunch. Breakfast was largely a standardised shake, so this partly reflects that product. Higher pre-meal glucose went with a smaller rise (−6.1 per 10 mg/dL), partly because starting high leaves less room to rise.
Glycemic status shifts the level
For the same food, people with T2D had responses 57.7 mg/dL × h larger than normoglycemic people (CI 34.5 to 80.9). Prediabetes was in between but not clearly different (16.7, CI −5.4 to 38.9). The carbohydrate slope was similar in the three groups (4.7, 4.0, 6.0 per 10 g), so status moved the level more than the steepness.
The same breakfast, three groups
The identical 712 kcal breakfast (66 g carbohydrate, 22 g protein, 42 g fat) eaten 88 times by 44 people gave a median iAUC of 43 (normoglycemic), 79 (prediabetes) and 114 (T2D) mg/dL × h, with single responses from 2 to 348.
Differences between people
Before adjustment, 19% of the variation in response lay between people (ICC 0.19); after adjustment the remaining participant-level share was 35%. Fixed effects explained 28% of the variance, 53% once each person's own average level was added. Person-specific carbohydrate slopes ranged from 0.1 to 13.9 per 10 g (median 5.0), all positive.
Predicting a new person (out-of-sample R²)
Bar length shows R² on a 0 to 50% scale. Teal = models that know the person.
Carbohydrate only
5%
Four macros
11%
+ pre-meal glucose, hour
11%
+ HbA1c, BMI, age
22%
Random forest, same inputs
27%
Predicting a new person is hard
Carbohydrate alone explained 5% of a held-out person's meals; all four macros 11%; adding pre-meal glucose and hour changed little (11%). Adding HbA1c, BMI and age doubled it to 22%, and a random forest reached 27% (CI 9 to 41%).
A person's own history helps
Within a person, a model fitted on their other days (carbohydrate and fiber only) explained 38% of held-out meals, versus 10% for the population model and 17% for just predicting the person's average. 17 of the 38 points come from knowing their typical level.
Within one person (held-out days, R²)
Bar length shows R² on a 0 to 50% scale. Test is optimistic: it trains on days after the test day.
Predict the person's average
17%
Population model
10%
Population + person's residual
32%
Person's own other days
38%
Participant 23
The person with the most analysable meals (35): a 55-year-old man with prediabetes. His carbohydrate slope (5.6 per 10 g) was close to the cohort's, but carbohydrate explained only 11% of his variation and his responses were largest at lunch and dinner, the opposite of the cohort.
Click an answer to reveal feedback. Each question locks after answering; options reshuffle on reset. Quizzes follow the report sections.
— / 54Overall quiz score (correct / answered, of 54)
✓ iAUC is the incremental area: only the part above the baseline counts, over 120 minutes, in mg/dL × hours. Peak rise is the highest value minus baseline.
Not quite. iAUC is the incremental area: only the part above the baseline counts, over 120 minutes, in mg/dL × hours. Peak rise is the highest value minus baseline.
Vocabulary · question 2 of 8
The pre-meal baseline is defined as:
✓ Baseline = mean Dexcom glucose in the 15 minutes before eating. It matters because the starting level limits how far glucose can rise.
Not quite. Baseline = mean Dexcom glucose in the 15 minutes before eating. It matters because the starting level limits how far glucose can rise.
Vocabulary · question 3 of 8
An ICC of 0.19 in the empty model means:
✓ The ICC is the between-person share of the variance. Most variation (about 81%) is between meals of the same person.
Not quite. The ICC is the between-person share of the variance. Most variation (about 81%) is between meals of the same person.
Vocabulary · question 4 of 8
Why give every participant a random intercept?
✓ Meals from one person share that person's physiology. A random intercept accounts for it so 1,173 meals are not treated like 1,173 independent people.
Not quite. Meals from one person share that person's physiology. A random intercept accounts for it so 1,173 meals are not treated like 1,173 independent people.
Vocabulary · question 5 of 8
An out-of-sample R² of 0.05 for the carbohydrate-only model means:
✓ R² is the share of variation explained. Out-of-sample means measured on held-out people, so it shows how well the model would do for a new person.
Not quite. R² is the share of variation explained. Out-of-sample means measured on held-out people, so it shows how well the model would do for a new person.
Vocabulary · question 6 of 8
What does 'isolated meal' mean?
✓ Isolation lets a glucose response be attributed to one meal. 322 meals were dropped because another meal overlapped.
Not quite. Isolation lets a glucose response be attributed to one meal. 322 meals were dropped because another meal overlapped.
Vocabulary · question 7 of 8
Which pairing of HbA1c thresholds and groups is correct?
✓ The report uses HbA1c below 5.7%, 5.7 to 6.4%, and 6.5% or higher, giving 15, 16 and 14 people.
Not quite. The report uses HbA1c below 5.7%, 5.7 to 6.4%, and 6.5% or higher, giving 15, 16 and 14 people.
Vocabulary · question 8 of 8
Which statement matches the Spearman correlation?
✓ Spearman works on ranks, so it handles skewed data such as iAUC and fits any consistently rising or falling pattern.
Not quite. Spearman works on ranks, so it handles skewed data such as iAUC and fits any consistently rising or falling pattern.
Section 1: Descriptives and cleaning
Section score: — / 6
Section 1 · question 1 of 6
How many people are in the dataset, and how do they split by HbA1c?
✓ 45 adults: 15 normoglycemic, 16 prediabetes and 14 with type 2 diabetes.
Not quite. 45 adults: 15 normoglycemic, 16 prediabetes and 14 with type 2 diabetes.
Section 1 · question 2 of 6
Which cleaning step removed the most meals?
✓ 1,706 → minus 174 implausible → minus 37 poor coverage → minus 322 not isolated = 1,173. Isolation removed the most.
Not quite. 1,706 → minus 174 implausible → minus 37 poor coverage → minus 322 not isolated = 1,173. Isolation removed the most.
Section 1 · question 3 of 6
Which of these is a nutrition plausibility rule?
✓ Plausibility rules are: under 50 kcal, a macro above 176 g, fiber above carbohydrate, or calories more than 40% below or 60% above what the macros imply. Coverage and isolation are separate steps.
Not quite. Plausibility rules are: under 50 kcal, a macro above 176 g, fiber above carbohydrate, or calories more than 40% below or 60% above what the macros imply. Coverage and isolation are separate steps.
Section 1 · question 4 of 6
Why was activity (METs) not used?
✓ The 45 files come in eight column layouts, so some variables exist for only some people. That is why activity was left out.
Not quite. The 45 files come in eight column layouts, so some variables exist for only some people. That is why activity was left out.
Section 1 · question 5 of 6
Where are most of the missing Dexcom minutes?
✓ About 90% of gaps are at the end, 5% are a warm-up at the start (median 70 minutes) and 5% are inside the recording.
Not quite. About 90% of gaps are at the end, 5% are a warm-up at the start (median 70 minutes) and 5% are inside the recording.
Section 1 · question 6 of 6
Why were Dexcom readings used for every glucose feature?
✓ Dexcom was more complete and finer-grained. Libre read about 32 mg/dL lower on average, so using both would have mixed two scales.
Not quite. Dexcom was more complete and finer-grained. Libre read about 32 mg/dL lower on average, so using both would have mixed two scales.
Section 2: Features and why
Section score: — / 7
Section 2 · question 1 of 7
Why is pre-meal baseline controlled for?
✓ A high starting level leaves less room to rise. This is also why the adjusted model shows a smaller rise at higher baseline.
Not quite. A high starting level leaves less room to rise. This is also why the adjusted model shows a smaller rise at higher baseline.
Section 2 · question 2 of 7
Why compute peak rise as well as iAUC?
✓ They answer slightly different questions. In this data they correlate 0.95, so peak rise gave the same pattern of results.
Not quite. They answer slightly different questions. In this data they correlate 0.95, so peak rise gave the same pattern of results.
Section 2 · question 3 of 7
Why was fiber divided by carbohydrate?
✓ Fiber per gram of carbohydrate had the strongest correlation with iAUC (−0.31).
Not quite. Fiber per gram of carbohydrate had the strongest correlation with iAUC (−0.31).
Section 2 · question 4 of 7
Which pair lists ONE per-person feature and ONE per-meal feature?
✓ Mean glucose, SD, CV and time in range are per person over the whole recording. Baseline, iAUC and peak rise are per meal.
Not quite. Mean glucose, SD, CV and time in range are per person over the whole recording. Baseline, iAUC and peak rise are per meal.
Section 2 · question 5 of 7
Why report glucose SD or CV, not only mean glucose?
✓ Level and swing are different things. The groups differ in both: mean 122, 135, 166 and SD 20.8, 26.7, 40.4.
Not quite. Level and swing are different things. The groups differ in both: mean 122, 135, 166 and SD 20.8, 26.7, 40.4.
Section 2 · question 6 of 7
In iAUC, what happens when glucose falls below the baseline?
✓ iAUC counts only the area above baseline, over the 2 hours after the meal (mg/dL × hours).
Not quite. iAUC counts only the area above baseline, over the 2 hours after the meal (mg/dL × hours).
Section 2 · question 7 of 7
Why were meal type and hour of day included?
✓ Breakfasts produced 24.0 mg/dL × h more than lunches, but breakfast was largely a standardised shake, so the effect may partly reflect that product.
Not quite. Breakfasts produced 24.0 mg/dL × h more than lunches, but breakfast was largely a standardised shake, so the effect may partly reflect that product.
Section 3a: Correlations
Section score: — / 6
Section 3a · question 1 of 6
Which meal feature had the strongest correlation with iAUC?
✓ Fiber per gram of carbohydrate (−0.31) was strongest, followed by carbohydrate (0.29) and the carbohydrate share of energy (0.25).
Not quite. Fiber per gram of carbohydrate (−0.31) was strongest, followed by carbohydrate (0.29) and the carbohydrate share of energy (0.25).
Section 3a · question 2 of 6
Protein (−0.05) and fat (0.05) correlated with iAUC. How should this be read?
✓ Both are close to zero. The mixed model later gave protein a small negative effect and fat no clear effect (CI includes 0).
Not quite. Both are close to zero. The mixed model later gave protein a small negative effect and fat no clear effect (CI includes 0).
Section 3a · question 3 of 6
Pre-meal glucose correlated −0.02 with iAUC, yet the adjusted model gave −6.1 per 10 mg/dL. What is a sensible explanation?
✓ This is the logic of adjustment (an interpretation, not stated as a result in the report). In the raw pooled data, group differences work against the within-group effect.
Not quite. This is the logic of adjustment (an interpretation, not stated as a result in the report). In the raw pooled data, group differences work against the within-group effect.
Section 3a · question 4 of 6
What is the purpose of the within-person correlation (value minus that person's average)?
✓ Carbohydrate was 0.32 and fiber −0.25 within a person, so the pattern was not just people who eat more carbs also spiking more.
Not quite. Carbohydrate was 0.32 and fiber −0.25 within a person, so the pattern was not just people who eat more carbs also spiking more.
Section 3a · question 5 of 6
Calories track carbohydrate (0.79) and fat (0.80). What follows for the analysis?
✓ Because nutrients are correlated, single correlations are misleading. The model estimates each nutrient while holding the others fixed.
Not quite. Because nutrients are correlated, single correlations are misleading. The model estimates each nutrient while holding the others fixed.
Section 3a · question 6 of 6
HbA1c correlated 0.75 with mean glucose across people (n = 45). Why does n = 45 matter?
✓ Between-person correlations use one value per person. Age (0.23, p = 0.13) and BMI (0.14, p = 0.36) were not clearly related to meal response.
Not quite. Between-person correlations use one value per person. Age (0.23, p = 0.13) and BMI (0.14, p = 0.36) were not clearly related to meal response.
Section 3b: Mixed-model results
Section score: — / 8
Section 3b · question 1 of 8
Carbohydrate: +6.0 mg/dL × h per 10 g (CI 5.0 to 7.0). What does this say?
✓ The unit is mg/dL × hours of area, not mg/dL. It is an average association, and person-specific slopes ranged from 0.1 to 13.9.
Not quite. The unit is mg/dL × hours of area, not mg/dL. It is an average association, and person-specific slopes ranged from 0.1 to 13.9.
Section 3b · question 2 of 8
The fiber coefficient is −15.5 per 10 g. How should it be read?
✓ Fiber is partly contained in the carbohydrate figure. The report advises not quoting it as a dose.
Not quite. Fiber is partly contained in the carbohydrate figure. The report advises not quoting it as a dose.
Section 3b · question 3 of 8
Fat: −1.6 per 10 g (CI −3.5 to 0.4). The right conclusion is:
✓ When a 95% interval spans zero, the data are compatible with a small negative effect, none, or a small positive one.
Not quite. When a 95% interval spans zero, the data are compatible with a small negative effect, none, or a small positive one.
Section 3b · question 4 of 8
T2D vs normoglycemic: +57.7 (CI 34.5 to 80.9). Prediabetes: +16.7 (CI −5.4 to 38.9). What does this show?
✓ Group differences are for the same food, since composition is adjusted for. The prediabetes interval includes 0.
Not quite. Group differences are for the same food, since composition is adjusted for. The prediabetes interval includes 0.
Section 3b · question 5 of 8
Carbohydrate slopes were 4.7, 4.0 and 6.0 per 10 g across the three groups. Conclusion?
✓ The T2D minus normoglycemic slope difference was 1.5 (CI −0.7 to 3.6), which includes 0.
Not quite. The T2D minus normoglycemic slope difference was 1.5 (CI −0.7 to 3.6), which includes 0.
Section 3b · question 6 of 8
Marginal R² was 28% and conditional R² was 53%. What is the gap?
✓ Fixed effects alone explain 28%; adding the person-level random intercepts brings it to 53%.
Not quite. Fixed effects alone explain 28%; adding the person-level random intercepts brings it to 53%.
Section 3b · question 7 of 8
The 3-hour-window carbohydrate effect was larger (9.6). Why is that expected?
✓ The area simply covers more time. The 2-hour effect was similar with overlapping meals included (5.4, 1,495 meals) and with Libre (6.3, 1,172 meals).
Not quite. The area simply covers more time. The 2-hour effect was similar with overlapping meals included (5.4, 1,495 meals) and with Libre (6.3, 1,172 meals).
Section 3b · question 8 of 8
Person-specific slopes ranged from 0.1 to 13.9, median 5.0, all positive. What does this tell us?
✓ All 45 slopes were positive, with a median of 5.0 per 10 g. Within a person, carbohydrate explained a median of 12% of the variation.
Not quite. All 45 slopes were positive, with a median of 5.0 per 10 g. Within a person, carbohydrate explained a median of 12% of the variation.
Section 3c: Prediction
Section score: — / 7
Section 3c · question 1 of 7
'Out-of-sample' in this report means:
✓ People were split into 9 groups; each group was held out once while the model was trained on the other 8.
Not quite. People were split into 9 groups; each group was held out once while the model was trained on the other 8.
Section 3c · question 2 of 7
Adding pre-meal glucose and hour to the four macros changed R² from 11% to 11%. What does it suggest?
✓ Adding the person's HbA1c, BMI and age doubled R² to 22%, so who the person is mattered far more than meal context.
Not quite. Adding the person's HbA1c, BMI and age doubled R² to 22%, so who the person is mattered far more than meal context.
Section 3c · question 3 of 7
The random forest reached 27% (CI 9 to 41%); the linear model with person traits 22% (CI 12 to 29%). What is a fair reading?
✓ The forest's interval is wide and contains the linear model's estimate. Don't over-read the difference.
Not quite. The forest's interval is wide and contains the linear model's estimate. Don't over-read the difference.
Section 3c · question 4 of 7
Why was the 95% interval made by resampling people, not meals?
✓ Resampling meals would make the interval too narrow, because one person's meals are correlated.
Not quite. Resampling meals would make the interval too narrow, because one person's meals are correlated.
Section 3c · question 5 of 7
Within a person, own-days model 38%, population model 10%, person's average 17%, hybrid 32%. How much of the 38% comes from knowing the typical level alone?
✓ Simply predicting the person's average already gives 17%, so the extra gain from the carbohydrate and fiber model on their own days is about 21 points.
Not quite. Simply predicting the person's average already gives 17%, so the extra gain from the carbohydrate and fiber model on their own days is about 21 points.
Section 3c · question 6 of 7
Why does the report call the personalised test optimistic?
✓ A real system would only know the past. The next step is forward-in-time splits.
Not quite. A real system would only know the past. The next step is forward-in-time splits.
Section 3c · question 7 of 7
Why report out-of-sample rather than in-sample R²?
✓ In-sample R² rewards models for memorising the data. The question that matters for a biomarker is how well it works on someone new.
Not quite. In-sample R² rewards models for memorising the data. The question that matters for a biomarker is how well it works on someone new.
Section 4: Interpretation and limitations
Section score: — / 7
Section 4 · question 1 of 7
The report links to Zeevi et al. (2015) and Berry et al. (2020). What is the shared message?
✓ The cohort result (11% explained by composition for a new person) agrees with this literature.
Not quite. The cohort result (11% explained by composition for a new person) agrees with this literature.
Section 4 · question 2 of 7
Using Coravos et al. (2019), the report says it supports only:
✓ It does not show that the measure predicts a health outcome (clinical validation) or that acting on it helps (utility).
Not quite. It does not show that the measure predicts a health outcome (clinical validation) or that acting on it helps (utility).
Section 4 · question 3 of 7
Why is 'this food spikes you' a premature claim?
✓ The report states the claim is premature because only analytical validation has been addressed.
Not quite. The report states the claim is premature because only analytical validation has been addressed.
Section 4 · question 4 of 7
What does the standard-breakfast result suggest?
✓ Medians of 43, 79 and 114 mg/dL × h separated the groups. Testing how well it classifies individuals against HbA1c is the next step.
Not quite. Medians of 43, 79 and 114 mg/dL × h separated the groups. Testing how well it classifies individuals against HbA1c is the next step.
Section 4 · question 5 of 7
How does noisy macronutrient logging affect the carbohydrate estimate?
✓ Measurement error in an exposure typically attenuates its estimated effect. The fiber coefficient is also fragile for similar reasons.
Not quite. Measurement error in an exposure typically attenuates its estimated effect. The fiber coefficient is also fragile for similar reasons.
Section 4 · question 6 of 7
The two sensors disagree (Libre 32 mg/dL lower, median per-person correlation 0.82). Why does iAUC help?
✓ Because iAUC measures the rise above each meal's own baseline, a constant offset between sensors cancels out. Differences in sensor dynamics do not.
Not quite. Because iAUC measures the rise above each meal's own baseline, a constant offset between sensors cancels out. Differences in sensor dynamics do not.
Section 4 · question 7 of 7
Which limitation affects what the findings describe?
✓ Results were similar with overlapping meals included (carbohydrate 5.4 per 10 g, 1,495 meals), but the conclusions still concern discrete meals.
Not quite. Results were similar with overlapping meals included (carbohydrate 5.4 per 10 g, 1,495 meals), but the conclusions still concern discrete meals.
Methods in practice: R and reproducibility
Section score: — / 5
Methods in practice: R and reproducibility · question 1 of 5
Why were column types fixed in read_csv?
✓ Fixing the types up front keeps the nutrition values intact when the 45 files are stacked.
Not quite. Fixing the types up front keeps the nutrition values intact when the 45 files are stacked.
Methods in practice: R and reproducibility · question 2 of 5
What are the 'EXPECT' comments in the R scripts for?
✓ They are a check on the R output. Small differences in the third decimal can be fine; a different count or sign is not.
Not quite. They are a check on the R output. Small differences in the third decimal can be fine; a different count or sign is not.
Methods in practice: R and reproducibility · question 3 of 5
In the mixed model, what does (1 | sid) mean?
✓ It is the random-intercept term: every participant (sid) has their own baseline level of response.
Not quite. It is the random-intercept term: every participant (sid) has their own baseline level of response.
Methods in practice: R and reproducibility · question 4 of 5
How were the linear-model out-of-sample R² values made to match the Python values in R?
✓ The fold assignment was hard-coded in the R script. The bootstrap intervals and the random forest depend on random draws and may differ slightly.
Not quite. The fold assignment was hard-coded in the R script. The bootstrap intervals and the random forest depend on random draws and may differ slightly.
Methods in practice: R and reproducibility · question 5 of 5
Which results in the report were NOT reproduced in R?
✓ Those came from the Python analysis only. Disclose this in your AI-use statement or reproduce them in R.
Not quite. Those came from the Python analysis only. Disclose this in your AI-use statement or reproduce them in R.
— / 54Overall quiz score
Core thesis
Meal composition has a real, plausible effect on the glucose rise, but it explains little of what a new person will do. Who the person is, and their own history, explain far more.
📊
Carbohydrate up, fiber down
Adjusted for pre-meal glucose, group and meal type, each extra 10 g of carbohydrate raised the 2-hour iAUC by about 6 mg/dL × h and each extra 10 g of fiber lowered it by about 15.
🧍
People differ more than meals
Composition, group, baseline and meal type together explained 28% of the variance; adding each person's own level brought it to 53%. Person-specific carbohydrate slopes ranged from 0.1 to 13.9.
🎯
Composition alone predicts a new person poorly
Out-of-sample R² was 11% for four macros and 22% with HbA1c, BMI and age. Knowing the person matters more than knowing the meal.
🔁
Personal history helps, partly through the average
A model on a person's own other days explained 38%, against 10% for the population model, but 17 points come from simply knowing their typical level.
⚖️
Only analytical validation
The measure behaves sensibly and separates groups (standard breakfast: 43, 79, 114). Clinical validation and utility are untested, so food-specific claims are premature.
🧪
Know what is in R and what is not
Descriptives, features, correlations, mixed models, the forest plot and out-of-sample R² were reproduced in R; leave-one-day-out and some secondary results were Python-only.
Clause’s comments: 98% of Start-ups fail. Why? This has been studied by Estell Discussion brought up topics No tool, just my thoughts (-; Build the winning team (without too high cash burn) Finding/selecting/addressing a significant medical problem/gap Be clear about the monetization strategy /monetisation model and verify Proven medical benefit. Reach regulatory approval Achieve UX…
Implementation Science: What Is It? — GIVEMEA Study Guide GIVEMEA Study Guide · Implementation Science Implementation Science: What Is It and Why Should I Care? Mark S. Bauer & JoAnn Kirchner · Psychiatry Research · 2019 Review Article Introductory / Conceptual Mental Health Context Special Issue Overview Overview Key Terms Methods References Quiz Takeaways 17–20Years…
Safavi et al. 2019 — GIVEMEA Study Guide GIVEMEA Study Guide · Digital Health Innovation Top-Funded Digital Health Companies & High-Burden Conditions Safavi, Mathews, Bates, Dorsey & Cohen · Health Affairs 38(1):115–123 · January 2019 Cross-Sectional Observational Secondary Research Top 20 Private US Companies Data through April 2017 Overview Key Terms Methods References Quiz Takeaways…
TAMI Business Model — Lighthouse Tech GIVEMEA Digital Health Intervention · Module 1 · Business Model Analysis TAMI Business Model Lighthouse Technology · lighthousetech.ch · Mendrisio, Switzerland ● Customer ● Value Proposition ● Value Chain ● Revenue Who is the customer? Customer Segments & Stakeholders Customers for the TAMI include individuals who are blind and those with low vision,…
Nahum-Shani et al. 2018 — JITAI Components & Design — GIVEMEA Study Guide Digital Health Interventions · mHealth · Intervention Design JITAIs in Mobile Health: Key Components & Design Principles Nahum-Shani et al. · Annals of Behavioral Medicine, 2018, Vol. 52 · pp. 446–462 Journal Article Ann Behav Med 2018 DOI 10.1007/s12160-016-9830-8 Conceptual / Review…