Preparing for the November 6th Presentation

Albuminuria Digital Biomarker Quiz – GIVEMEA Study Guide
GIVEMEA Study Guide · Digital Biomarkers, Kidney Dysfunction Project

Can a Digital Biomarker Identify Albuminuria? A Four-Level Quiz on Our Results, Methods and Interpretation

Group project, CAS Digital Health 2026/27 · AI-READI data · Presentation 6 November 2026

Cross-sectional Logistic regression 465 adults, 77 cases 20 questions, 4 levels
465Analytic sample
77With albuminuria
0.65First model AUROC
0.546CV baseline (M1)
0.589CV full model (M3)
Central Finding
Our model separates people with and without albuminuria only modestly. We cannot show that wearable or context features improve it, and time in range is the only possible lead.

How to use this quiz

Work through the levels in order. Each level builds on the last: recall, then method, then interpretation, then critical judgement. Flip through Key Terms first if you want a warm-up. The score runs across all 20 questions, and options reshuffle each time you reset.

What the project asked

Can clinical, CGM, wearable and self-reported features identify albuminuria (UACR of 30 mg/g or more) in adults with prediabetes or diabetes? We frame this as a diagnostic biomarker, because our data are a single snapshot in time.

The difficulty ladder

LevelWhat it testsQuestions
Level 1 · FoundationsRecall the question, the sample and the headline numbers.5
Level 2 · MethodsUnderstand why the analysis was built the way it was.5
Level 3 · InterpretationRead the numbers correctly and avoid common misreadings.5
Level 4 · Critical thinkingJudge the limits, the framing and what we can honestly claim.5

■ Our results   ■ Methods   ■ Interpretation traps   ■ Biomarker framework

Our results

UACR and albuminuria
tap to define
UACR and albuminuria
UACR is the urine albumin-to-creatinine ratio. Albuminuria here means UACR of 30 mg/g or more, measured at one baseline visit. It is one of the earliest measurable signs of kidney damage in diabetes.
AUROC
tap to define
AUROC
Area under the ROC curve. 1.0 is perfect and 0.50 is a coin toss. 0.65 means a random person with albuminuria is ranked above a random person without it about 65% of the time.
Time in range (TIR)
tap to define
Time in range (TIR)
Percentage of CGM readings between 70 and 180 mg/dL. In our model it carried the largest SHAP share (36.6%) and is the only possible lead (+0.049, p = 0.108).
Analytic sample
tap to define
Analytic sample
465 adults from 1,052: minus 365 healthy, minus 42 missing outcome or HbA1c, minus 180 missing a feature. 77 (16.6%) have albuminuria.

Methods

Nested models
tap to define
Nested models
Each model holds the previous features plus new ones: M1 (age, sex, BMI, HbA1c), M2 (M1 + time in range), M3 (M2 + sleep regularity, peak step rate, walking infrastructure, food insecurity).
Repeated stratified 5-fold CV
tap to define
Repeated stratified 5-fold CV
Data are split into five folds, each used once as the test set, keeping the case proportion similar. We repeated this 20 times so all 77 cases are tested.
Bootstrap 95% CI
tap to define
Bootstrap 95% CI
A confidence interval from 2,000 resamples. Ours use predictions averaged across repeats, so they are probably slightly too narrow.
DeLong test
tap to define
DeLong test
A test comparing two AUROCs on the same people. Applied to the averaged cross-validated predictions it is approximate.
Events per predictor
tap to define
Events per predictor
Rule of thumb: at least 10 cases per predictor. We have about 9 (61/7 in the first model, 77/9 in the cross-validated full model), with about 7 per fold.

Interpretation traps

No measurable gain
tap to define
No measurable gain
Wording for M3 versus M2 (-0.005, CI -0.056 to +0.048): no evidence of a gain, but a small benefit cannot be ruled out. Not the same as 'adds nothing'.
SHAP versus added value
tap to define
SHAP versus added value
SHAP shows how much a model leans on each feature. Added value is whether prediction is better with the feature than without. Wearables took about a quarter of SHAP yet gave no measurable gain.
Reverse causation
tap to define
Reverse causation
In cross-sectional data we cannot tell whether behaviour came before kidney damage. Poorer kidney health might reduce activity.
Internal versus external validation
tap to define
Internal versus external validation
We validated only inside one dataset (hold-out split and cross-validation). No second cohort, so real-world performance is likely the same or worse.

Biomarker framework

Digital biomarker
tap to define
Digital biomarker
A measure from sensors and algorithms, often outside the clinic. Coravos et al. (2019) separate the sensor layer from the algorithm layer.
Verification, validation, fit for purpose
tap to define
Verification, validation, fit for purpose
Coravos's three steps: bench-test the sensor, validate against the clinical outcome in the target group, and show it suits its intended use. We did only part of the validation step.
Diagnostic versus prognostic
tap to define
Diagnostic versus prognostic
Diagnostic detects what is present now. Prognostic predicts what will happen. Our single snapshot fits diagnostic or screening. The label is our own inference.
Peak step rate
tap to define
Peak step rate
Average of each day's highest step rate in one recording interval. Garmin intervals vary in length, and very short ones inflate it, so we use it as a rough, relative measure.

Level 1 · Foundations

Recall the question, the sample and the headline numbers. Click an answer to reveal feedback. Each question locks after answering.

Level 1 · Foundations · Question 1 of 20
What outcome did our model try to identify?
✓ Correct. The outcome was alb_binary: UACR of 30 mg/g or more versus below 30, measured once. The clinical rule that albuminuria must persist for 3 months could not be applied.
Not quite. The outcome was UACR of 30 mg/g or more at a single baseline visit. It was not a change in eGFR, not the severe threshold of 300 mg/g, and not a CKD stage.
Level 1 · Foundations · Question 2 of 20
How many people were in the final analytic sample, and how many had albuminuria?
✓ Correct. The sample went from 1,052 to 687 (healthy excluded), to 645 (missing outcome or HbA1c), to 465 (missing any of the seven features). Of the 465, 77 (16.6%) had albuminuria and 388 did not.
Not quite. 1,052 is everyone in the course extract and 645 is the count before dropping people with a missing feature. The final sample was 465, of whom 77 (16.6%) had albuminuria and 388 did not.
Level 1 · Foundations · Question 3 of 20
The first model had an AUROC of 0.65 (95% CI 0.49 to 0.81). Which statement is correct?
✓ Correct. An AUROC of 0.65 means a randomly chosen person with albuminuria is ranked above one without about 65% of the time. Because the interval reaches 0.49, a coin toss (0.50) is still plausible. Only 16 test cases made the interval wide.
Not quite. AUROC is a ranking measure, not the percentage of correct predictions. The point estimate of 0.65 is above 0.50, but the interval (0.49 to 0.81) includes 0.50, so we cannot rule out chance.
Level 1 · Foundations · Question 4 of 20
In the first model's SHAP summary, which single feature carried the largest share of the attribution?
✓ Correct. Time in range carried 36.6%, ahead of HbA1c (19.6%) and sleep regularity (16.6%). Glycaemic features together carried 56.2%.
Not quite. Time in range was the largest at 36.6%, followed by HbA1c (19.6%), sleep regularity (16.6%) and peak step rate (8.9%).
Level 1 · Foundations · Question 5 of 20
Why were the 365 healthy participants excluded from the analysis?
✓ Correct. Including healthy people would let the model learn something close to prediabetes or diabetes status, not albuminuria within people at risk. We restricted the question to people with prediabetes or diabetes.
Not quite. The reason was about the question we wanted to answer: with healthy people in the sample, glycaemic status alone would largely separate the groups and inflate performance.
– / 20Running score across all levels

Level 2 · Methods

Understand why the analysis was built the way it was. Click an answer to reveal feedback. Each question locks after answering.

Level 2 · Methods · Question 6 of 20
Why was the first model's classification threshold lowered from 0.50 to 0.15?
✓ Correct. With a low prevalence, a 0.50 cut-off would flag almost nobody. At 0.15 the model found 9 of 16 cases (sensitivity 0.56) and ruled out 52 of 78 non-cases (specificity 0.67), with 26 false positives and 7 missed cases. The threshold does not change the AUROC.
Not quite. The AUROC does not depend on the threshold. The threshold was lowered because only about 17% of people have albuminuria, so 0.50 would flag very few. A lower threshold usually adds false positives, not removes them.
Level 2 · Methods · Question 7 of 20
What is the main advantage of repeated stratified 5-fold cross-validation over the single 80/20 split?
✓ Correct. The single split tested on only 16 cases. Cross-validation tests on all 77 cases across folds, and stratifying keeps the proportion of cases similar in each fold. It is still internal validation on one dataset.
Not quite. Cross-validation does not raise the AUROC, remove uncertainty or count as external validation. Its advantage is that all 77 cases are used as test cases, which makes the estimate more stable.
Level 2 · Methods · Question 8 of 20
In the cross-validated models, age and sex were put in the baseline (M1) and time in range was left out of it. Why?
✓ Correct. A digital biomarker is only worth adding if it improves on what any clinic already knows. Step rate falls with age, so without age in the model it could act as a proxy. Time in range is a CGM-derived measure, so it was added as a separate step to test its value.
Not quite. The reasons were about fair testing: age and sex are routinely available and can act as hidden proxies for sensor features, and time in range is itself digital, so it was added as its own step.
Level 2 · Methods · Question 9 of 20
After the course professor's reply, how should we describe 'mean peak step rate'?
✓ Correct. Garmin intervals have no minimum length, so a few steps in a very short interval look like a very high per-minute rate. That is why we see a mean near 212 and a maximum of 862. The professor said it is fine to use as a relative measure with a brief note that it is imprecise.
Not quite. It is not a day-long or 30-minute cadence. It is the average of each day's single highest interval rate, and short recording intervals inflate it, so we treat it as a rough, relative measure.
Level 2 · Methods · Question 10 of 20
We compared three nested models. What does 'nested' mean here?
✓ Correct. Because each model adds features to the last, the AUROC difference between neighbouring models isolates what that block of features adds. All three were run on the same 465 people.
Not quite. 'Nested' means each model includes the previous model's features and adds more, so the difference in AUROC shows what the added features contribute.
– / 20Running score across all levels

Level 3 · Interpretation

Read the numbers correctly and avoid common misreadings. Click an answer to reveal feedback. Each question locks after answering.

Level 3 · Interpretation · Question 11 of 20
Adding time in range to the baseline changed the AUROC by +0.049 (95% CI -0.010 to +0.109, p = 0.108). What is the best reading?
✓ Correct. The estimate is positive and about 5 AUROC points, but the interval runs from slightly below zero to +0.11 and p = 0.108. So time in range is the only possible lead, and even that is not statistically clear.
Not quite. The interval includes zero and p is above 0.05, so the gain is not statistically clear. A positive estimate alone is not proof, and nothing here suggests harm.
Level 3 · Interpretation · Question 12 of 20
Adding sleep regularity, peak step rate, walking infrastructure and food insecurity changed the AUROC by -0.005 (95% CI -0.056 to +0.048). Which sentence is best?
✓ Correct. The change is half a percentage point and the interval includes zero, so there is no evidence of a gain. The upper end reaches about +0.05, so a small benefit is not excluded. The result does not say the wearables do not work, and it does not say they do.
Not quite. 'Adds nothing' is too strong, because the interval reaches about +0.05. 'Clearly improves' and 'got worse' both over-read a difference that is far smaller than the noise. The careful wording is 'no measurable gain; a small benefit cannot be ruled out'.
Level 3 · Interpretation · Question 13 of 20
Wearable features carried about a quarter (25.5%) of the first model's SHAP attribution, yet adding them did not measurably improve cross-validated AUROC. How do these fit together?
✓ Correct. Think of a student who uses a calculator on every question: it is clearly used, but it only helps if the score is higher with it. SHAP describes use; the nested comparison tests added value (0.589 with, 0.595 without). Also, the SHAP numbers come from the first model, which has no age or sex.
Not quite. The two analyses answer different questions and do not contradict each other. SHAP shows reliance within one model. Added value is the with-versus-without comparison, and SHAP says nothing about causation.
Level 3 · Interpretation · Question 14 of 20
The first model had 61 training cases for 7 predictors. What does that imply?
✓ Correct. 61 divided by 7 is about 9. The cross-validated full model has 9 predictors (seven features plus age and sex) and 77 cases, also about 9, and each fold fits on roughly 62 cases, closer to 7. These are slightly under the guideline of 10.
Not quite. The ratio is cases divided by predictors: 61 / 7 is about 9, a little under the guideline of 10. It is not 61 per predictor, and it is not far below the guideline.
Level 3 · Interpretation · Question 15 of 20
The baseline model M1 (age, sex, BMI, HbA1c) had a cross-validated AUROC of 0.546 (95% CI 0.465 to 0.616). What does the interval tell us?
✓ Correct. This is easy to miss. The baseline is not clearly above chance, time in range adds a gain whose interval includes zero, and the remaining features add no measurable gain. The full model's lower bound (0.51) just clears 0.50, but our bootstrap intervals are probably slightly too narrow.
Not quite. Because the interval includes 0.50, the baseline model cannot be told apart from a coin toss. That is why the story is a modest signal with no firm conclusions.
– / 20Running score across all levels

Level 4 · Critical thinking

Judge the limits, the framing and what we can honestly claim. Click an answer to reveal feedback. Each question locks after answering.

Level 4 · Critical thinking · Question 16 of 20
Why do we call our target a diagnostic (or screening) biomarker and not a prognostic one?
✓ Correct. A prognostic claim needs follow-up data. The label is our own inference from the design, not a statement in the Coravos paper, and the group still has to agree on one wording.
Not quite. The reason is the design: one snapshot in time supports a detection question. The category label is our inference, not something Coravos assigned, and prognosis would need longitudinal data.
Level 4 · Critical thinking · Question 17 of 20
What validation have we done, and what have we not done?
✓ Correct. Everything comes from one AI-READI course extract. Bootstrap intervals describe uncertainty within that dataset and are not external validation. In Coravos's terms we did part of the validation step, with modest results, and nothing on fitness for purpose.
Not quite. We have only internal validation on one dataset: a hold-out split and repeated cross-validation. Bootstrap resampling is not external validation, and we have no second cohort.
Level 4 · Critical thinking · Question 18 of 20
A reader proposes comparing our AUROC with the C-index of 0.783 reported by Xu et al. (2021). Why is that not a fair benchmark?
✓ Correct. Xu is useful for the direction of the HbA1c effect (OR 1.29 per +1% HbA1c, 213 people, about 26.7 months of follow-up) but is not a like-for-like benchmark for a cross-sectional, sensor-based model. Most of its discrimination comes from baseline ACR (C-index 0.736 alone).
Not quite. The problem is the design. Xu followed 213 people with type 2 diabetes over time, used baseline ACR and orosomucoid, selected variables stepwise and reported accuracy as well as the C-index. Our setting is cross-sectional and does not use ACR.
Level 4 · Critical thinking · Question 19 of 20
Hyder et al. (2026) review how social determinants of health shape chronic disease. What can we use it for?
✓ Correct. The abstract does not mention kidney disease or food insecurity by name, so it supports the reasoning for including the two context features and nothing more. The evidence for these two features is indirect or pathway-only.
Not quite. It is a narrative review used for framing. It does not show that food insecurity or walkability predicts or causes albuminuria.
Level 4 · Critical thinking · Question 20 of 20
Which one-sentence summary of our work is the most defensible?
✓ Correct. This keeps the effect size, the uncertainty, the design limits and the lead together. We describe it as a first exploratory step toward a diagnostic biomarker, not a clinical tool.
Not quite. The defensible version keeps the modest effect, the wide uncertainty, the cross-sectional and internal-only design, and the fact that only time in range is a possible lead. Claims of reliability, clinical readiness or proof overstate the evidence.
– / 20Running score across all levels
Core Thesis
A first exploratory step toward a diagnostic digital biomarker: a modest signal, no measurable gain from wearable and context features, internal validation only, and honest limits.
  • 📊

    A modest signal, honestly reported

    Cross-validated AUROC is about 0.55 to 0.60. The first model's 0.65 (CI 0.49 to 0.81) is consistent with that. The baseline model alone is not clearly above chance.

  • 🎯

    Time in range is the only possible lead

    Adding it moved AUROC by +0.049, but the interval includes zero (p = 0.108). Treat it as a hypothesis for a larger sample.

  • 💡

    No measurable gain from wearable and context features

    The change was -0.005 (CI -0.056 to +0.048). A small benefit cannot be ruled out, and the result does not say wearables do not work.

  • ⚙️

    SHAP describes use, not value

    About a quarter of the first model's attribution went to wearable features, but added value needs the with-versus-without comparison.

  • ⚖️

    Diagnostic framing, internal validation only

    Cross-sectional data support a detection question. Nothing here is externally validated, and it is a first exploratory step, not a clinical tool.

  • 🏛️

    Compare like with like

    Xu et al. (2021) predicts incident albuminuria with baseline ACR, so its C-index is not a benchmark for our AUROC.

Similar Posts

Leave a Reply