Accelerometry-Derived Digital Biomarkers for Cardiometabolic Risk: A Population-Representative Tabular Benchmark with Uncertainty Quantification

Federico Felizzi

arxiv: 2606.30702 · v1 · pith:BYROUHP4new · submitted 2026-06-29 · 💻 cs.LG · cs.AI· stat.ML

Accelerometry-Derived Digital Biomarkers for Cardiometabolic Risk: A Population-Representative Tabular Benchmark with Uncertainty Quantification

Federico Felizzi This is my paper

Pith reviewed 2026-07-01 07:14 UTC · model grok-4.3

classification 💻 cs.LG cs.AIstat.ML

keywords accelerometryNHANEStabular learningconformal predictioncardiometabolic biomarkersHbA1cC-reactive proteintriglycerides

0 comments

The pith

TabPFN v2 leads predictions of HbA1c and CRP from accelerometry data in an NHANES benchmark, yet triglycerides resist prediction and conformal intervals under-cover some demographic subgroups.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper builds a benchmark from NHANES 2003-2006 accelerometry and covariate records on 1,381 adults to forecast three cardiometabolic biomarkers. It pits ridge regression, XGBoost, and TabPFN v2 against one another on this population-representative tabular setting. TabPFN v2 records the highest R-squared scores for HbA1c and CRP while triglycerides stay near zero predictability. Split conformal prediction supplies 90 percent intervals that hit marginal targets for two outcomes but fall short inside subgroups such as Mexican American participants. The work therefore shows both the promise of foundation models for activity-derived biomarkers and the gap between marginal coverage guarantees and the conditional coverage needed for equitable use.

Core claim

TabPFN v2 achieves the best overall performance (HbA1c R²=0.156, CRP R²=0.383), while triglycerides remain largely unpredictable (R² < 0.05), consistent with known genetic dominance; marginal coverage aligns with the 90% target for CRP and HbA1c but falls below for triglycerides, with localized undercoverage e.g. for Mexican American participants.

What carries the argument

The NHANES Accelerometry Cardiometabolic Benchmark table of accelerometry phenotypes, lifestyle covariates, and laboratory biomarkers, paired with split conformal prediction to produce distribution-free 90% prediction intervals.

If this is right

TabPFN v2 should be preferred over ridge regression and XGBoost for these specific biomarker tasks on similar tabular inputs.
Activity and lifestyle features alone cannot reliably predict fasting triglycerides.
Marginal conformal intervals meet nominal coverage on average but do not guarantee coverage inside every demographic stratum.
Methods that enforce conditional coverage will be required before the intervals can support clinical decisions without fairness gaps.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Adding genetic variants as features might lift triglyceride predictability above the activity-only ceiling observed here.
Repeating the benchmark on later NHANES waves could reveal whether secular changes in activity levels alter the reported R-squared values.
Per-stratum recalibration of conformal scores offers one direct route to close the observed subgroup coverage shortfalls.
The same tabular setup could serve as a testbed for newer foundation models to measure progress on the hard triglyceride target.

Load-bearing premise

The processed NHANES 2003-2006 accelerometry and covariate table supplies a representative and sufficiently informative feature set for the three prediction tasks and for valid application of split conformal prediction.

What would settle it

An independent cohort with matched accelerometry, covariates, and the same three biomarkers that yields materially lower R-squared values for TabPFN v2 or coverage rates outside the reported margins would falsify the performance and calibration claims.

Figures

Figures reproduced from arXiv: 2606.30702 by Federico Felizzi.

**Figure 1.** Figure 1: Empirical conformal coverage (90% target) by demographic subgroup for TabPFN v2. Colour indicates coverage level: green ≥0.90, red <0.90. Why TabPFN v2 outperforms task-specific models. The superior performance of TabPFN v2 on a dataset of N=1,381 is consistent with its design objective: the model’s pretraining on synthetic datasets spanning diverse causal structures effectively encodes a broad prior ov… view at source ↗

read the original abstract

Structured tabular data dominates clinical medicine, yet existing benchmarks fail to reflect real-world properties like complex survey sampling, demographic oversampling, and subgroup fairness. We introduce the NHANES Accelerometry Cardiometabolic Benchmark, derived from NHANES 2003-2006, comprising 1,381 adults with hip-worn accelerometry, fasting laboratory biomarkers, dietary intake, and anthropometrics. We evaluate three tabular learning methods -- ridge regression, XGBoost, and the foundation model TabPFN v2 -- to predict glycated haemoglobin (HbA1c), fasting triglycerides, and C-reactive protein (CRP) from activity phenotypes and lifestyle covariates. TabPFN v2 achieves the best overall performance (HbA1c R^2=0.156, CRP R^2=0.383), while triglycerides remain largely unpredictable (R^2 < 0.05), consistent with known genetic dominance. We apply split conformal prediction to generate distribution-free 90% prediction intervals and evaluate demographic coverage equity across sex and race/ethnicity subgroups. Marginal coverage aligns with the 90% target for CRP and HbA1c but falls below for triglycerides. At the subgroup level, we observe localized undercoverage (e.g., HbA1c for Mexican American participants), illustrating the gap between marginal guarantees and the conditional coverage required for clinical fairness. Code and data are at https://github.com/felizzi/nhanes-accel-cardiometabolic-benchmark.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This paper supplies a new NHANES accelerometry benchmark for three cardiometabolic targets with TabPFN and conformal intervals, but the coverage claims rest on unadjusted complex survey sampling.

read the letter

The punchline is that the work creates a reusable tabular benchmark from NHANES 2003-2006 accelerometry data on 1,381 adults and shows TabPFN v2 leading on HbA1c and CRP while triglycerides stay near zero R-squared. It also checks 90% conformal coverage across demographic subgroups and releases the code and data.

What is actually new is the specific pairing of hip-worn activity phenotypes with those three lab targets inside a population-representative sample that includes oversampling and the explicit subgroup coverage check. Prior tabular health benchmarks rarely embed the survey design this way. The data release and the observation that marginal coverage holds for two targets but drops for triglycerides and some Mexican American subgroups are concrete and usable.

The soft spot is the split conformal procedure. NHANES relies on stratified multistage sampling with deliberate oversampling and survey weights, which breaks the exchangeability assumption required for the distribution-free coverage guarantee. The abstract applies the standard unweighted version without mentioning design adjustments or weighted conformal variants, so the reported coverage numbers and the fairness conclusions drawn from them need verification.

This is for people working on tabular clinical models or conformal methods who want a testbed that forces attention to sampling and subgroup issues. A reader building or evaluating health prediction pipelines would get value from the released table and the performance baselines.

It deserves peer review. The benchmark and data contribution are solid enough to justify referee time even if the conformal section requires fixes.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces the NHANES Accelerometry Cardiometabolic Benchmark derived from NHANES 2003-2006, comprising 1,381 adults with hip-worn accelerometry, fasting biomarkers (HbA1c, triglycerides, CRP), dietary intake, and anthropometrics. It benchmarks ridge regression, XGBoost, and TabPFN v2 for predicting the three biomarkers from activity phenotypes and covariates, reports TabPFN v2 as best overall (HbA1c R²=0.156, CRP R²=0.383, triglycerides R²<0.05), and applies split conformal prediction to produce 90% prediction intervals, finding marginal coverage near target for CRP and HbA1c but shortfalls for triglycerides and localized undercoverage in subgroups such as Mexican American participants for HbA1c.

Significance. If the results hold after addressing the sampling-design issue, the work supplies a population-representative tabular benchmark that incorporates real-world survey complexities and demographic oversampling, together with public code and data at the cited GitHub repository. The performance ordering and the explicit demonstration of the gap between marginal and subgroup coverage would be useful for tabular ML evaluation in clinical risk prediction.

major comments (2)

[Abstract and Methods (conformal prediction)] Abstract and Methods (split conformal prediction paragraph): standard split conformal is applied directly to the processed 1,381-row table without incorporating NHANES survey weights, stratification, or clustering adjustments. Because the design deliberately oversamples Mexican Americans, African Americans, and adults ≥60, the exchangeability assumption required for distribution-free marginal coverage is not satisfied; this directly threatens the reported 90% coverage figures and the subgroup undercoverage conclusions.
[Methods (data processing)] Methods (data-processing and phenotype-extraction subsection): the manuscript provides no explicit exclusion criteria, accelerometry-feature definitions, or verification that the final analytic sample remains representative after cleaning. These details are load-bearing for the claim that the table supplies a valid population-representative feature set for both the prediction tasks and the conformal procedure.

minor comments (1)

[Abstract] The abstract states that the benchmark 'reflects real-world properties like complex survey sampling' yet does not describe how (or whether) the sampling design enters the conformal procedure; a single clarifying sentence would resolve the apparent inconsistency.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on the handling of NHANES survey design in conformal prediction and the need for greater transparency in data processing. We address each major comment below.

read point-by-point responses

Referee: Abstract and Methods (conformal prediction) Abstract and Methods (split conformal prediction paragraph): standard split conformal is applied directly to the processed 1,381-row table without incorporating NHANES survey weights, stratification, or clustering adjustments. Because the design deliberately oversamples Mexican Americans, African Americans, and adults ≥60, the exchangeability assumption required for distribution-free marginal coverage is not satisfied; this directly threatens the reported 90% coverage figures and the subgroup undercoverage conclusions.

Authors: We agree that standard split conformal prediction was applied directly to the analytic sample without survey weights or design adjustments, and that the complex sampling (stratification, clustering, and oversampling) means the exchangeability assumption for marginal coverage may not hold with respect to the full target population. This is a genuine limitation for interpreting the coverage results as population-representative. We will revise the Methods and Discussion to explicitly acknowledge this, state that the reported coverage applies to the processed sample, and note that extending conformal methods to incorporate survey design is an open direction for future work. revision: yes
Referee: Methods (data processing) Methods (data-processing and phenotype-extraction subsection): the manuscript provides no explicit exclusion criteria, accelerometry-feature definitions, or verification that the final analytic sample remains representative after cleaning. These details are load-bearing for the claim that the table supplies a valid population-representative feature set for both the prediction tasks and the conformal procedure.

Authors: We will expand the data-processing subsection to include explicit exclusion criteria (minimum valid wear time, complete biomarker and covariate data, age restrictions), precise definitions of all accelerometry phenotypes (e.g., cutpoints and aggregation rules for MVPA, sedentary time, and total counts), and a table or text comparing key demographic proportions in the analytic sample versus the full eligible NHANES 2003-2006 population. These details will be added to support the representativeness claim. revision: yes

Circularity Check

0 steps flagged

No circularity: all claims are direct empirical evaluations on held-out data

full rationale

The paper reports R^2 values, coverage rates, and subgroup observations computed directly from model fits and conformal intervals on train/test splits of the processed NHANES table. No equations define a quantity in terms of itself, no fitted parameter is relabeled as a prediction, and no load-bearing premise rests on self-citation. The conformal coverage numbers are empirical frequencies, not quantities forced by internal definitions or ansatzes. The sampling-design validity concern is an external assumption issue, not a reduction of the reported results to the paper's own inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Empirical benchmark paper; no new theoretical constructs or fitted constants are introduced beyond standard ML practice.

axioms (1)

domain assumption Split conformal prediction yields valid marginal coverage under exchangeability of calibration and test points
Invoked to generate 90% prediction intervals and to interpret marginal vs. subgroup coverage

pith-pipeline@v0.9.1-grok · 5799 in / 1428 out tokens · 28802 ms · 2026-07-01T07:14:18.298521+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

10 extracted references · 10 canonical work pages · 1 internal anchor

[1]

Representation learning for tabular data: A comprehensive survey

Representation Learning for Tabular Data: A Comprehensive Survey , author=. arXiv preprint arXiv:2504.16109 , year=

work page arXiv
[2]

Advances in Neural Information Processing Systems , volume=

Revisiting Deep Learning Models for Tabular Data , author=. Advances in Neural Information Processing Systems , volume=

work page
[3]

Nature , volume=

Accurate predictions on small data with a tabular foundation model , author=. Nature , volume=

work page
[4]

A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification

A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification , author=. arXiv preprint arXiv:2107.07511 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[5]

Advances in Neural Information Processing Systems , volume=

When do neural nets outperform boosted trees on tabular data? , author=. Advances in Neural Information Processing Systems , volume=

work page
[6]

Chen, Tianqi and Guestrin, Carlos , booktitle=

work page
[7]

Physical activity in the

Troiano, Richard P and Berrigan, David and Dodd, Kevin W and M. Physical activity in the. Medicine & Science in Sports & Exercise , volume=

work page
[8]

Nature , volume=

Biological, clinical and population relevance of 95 loci for blood lipids , author=. Nature , volume=

work page
[9]

Archives of Internal Medicine , volume=

Glycemic control and coronary heart disease risk in persons with and without diabetes , author=. Archives of Internal Medicine , volume=

work page
[10]

The Annals of Statistics , volume=

Predictive inference with the jackknife+ , author=. The Annals of Statistics , volume=. 2021 , publisher=

work page 2021

[1] [1]

Representation learning for tabular data: A comprehensive survey

Representation Learning for Tabular Data: A Comprehensive Survey , author=. arXiv preprint arXiv:2504.16109 , year=

work page arXiv

[2] [2]

Advances in Neural Information Processing Systems , volume=

Revisiting Deep Learning Models for Tabular Data , author=. Advances in Neural Information Processing Systems , volume=

work page

[3] [3]

Nature , volume=

Accurate predictions on small data with a tabular foundation model , author=. Nature , volume=

work page

[4] [4]

A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification

A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification , author=. arXiv preprint arXiv:2107.07511 , year=

work page internal anchor Pith review Pith/arXiv arXiv

[5] [5]

Advances in Neural Information Processing Systems , volume=

When do neural nets outperform boosted trees on tabular data? , author=. Advances in Neural Information Processing Systems , volume=

work page

[6] [6]

Chen, Tianqi and Guestrin, Carlos , booktitle=

work page

[7] [7]

Physical activity in the

Troiano, Richard P and Berrigan, David and Dodd, Kevin W and M. Physical activity in the. Medicine & Science in Sports & Exercise , volume=

work page

[8] [8]

Nature , volume=

Biological, clinical and population relevance of 95 loci for blood lipids , author=. Nature , volume=

work page

[9] [9]

Archives of Internal Medicine , volume=

Glycemic control and coronary heart disease risk in persons with and without diabetes , author=. Archives of Internal Medicine , volume=

work page

[10] [10]

The Annals of Statistics , volume=

Predictive inference with the jackknife+ , author=. The Annals of Statistics , volume=. 2021 , publisher=

work page 2021