Working paper · draft

A Three-Tier Marker Framework with Asymmetric Compound Detection Rules for Metabolic Syndrome Screening

Methodology and Validation Against All-Cause and Cardiovascular Mortality in Pooled NHANES (2007–2016)

Dmitrii Shibakov · Zenlo LLC · preprint draft v3.0 · CC-BY 4.0

Preprint draft — not yet posted and not certified by peer review. This reports new methodology and analysis and should not be used to guide clinical practice.

Abstract

Background. Clinical screening rules differ in how they combine biomarkers. Flat any-abnormal logic flags a marker cluster when any single component is abnormal, whereas guideline syndrome definitions require a minimum count of criteria (e.g., metabolic syndrome, ≥3 of 5, NCEP ATP-III). We formalize a three-tier marker taxonomy (pathognomonic / confirmatory / supportive) mapped to asymmetric compound detection rules, and test whether the asymmetric ≥N-of-M rule captures mortality-relevant risk that flat any-trigger logic dilutes.

Methods. Deterministic detectors implementing guideline thresholds were applied to public NHANES data. For outcome validation, we pooled the fasting subsample of five continuous cycles (2007–2008 to 2015–2016; primary cohort age 20–64, N=9,245) and linked participants to the NCHS 2019 Public-Use Linked Mortality File (follow-up to 31 December 2019; 354 all-cause and 81 cardiovascular deaths). Using survey-weighted Cox proportional-hazards models (pooled fasting weights; Taylor-linearized variance with primary sampling units and strata), we compared a flat any-trigger rule (≥1 of 5 components) with the asymmetric ≥3-of-5 rule and with component count (0–5), adjusting for age, sex, and race/ethnicity. We additionally report a cross-sectional analysis (NHANES 2017–2018) of rule behavior and of implementation faithfulness relative to a large-language-model (LLM) screening layer.

Results. The flat any-trigger rule classified 80.4% of the cohort as positive but was not associated with all-cause mortality (HR 1.26, 95% CI 0.87–1.83). The asymmetric ≥3-of-5 rule classified 29.2% and was associated with both all-cause (HR 1.56, 1.22–1.98) and cardiovascular mortality (HR 2.37, 1.39–4.03). Risk rose graded with component burden, reaching significance at counts 4 (HR 2.07) and 5 (HR 2.43) rather than stepping at 3. Discrimination was similar between rules (Harrell's C 0.766 vs 0.769). Estimates were consistent unweighted (≥3-of-5 HR 1.40) and in the age ≥20 cohort (HR 1.15). Cross-sectionally, the flat rule flagged 75.1% of individuals who did not meet the syndrome definition, which the ≥3-of-5 rule reduced to the syndrome prevalence (32.3%). Deterministic detectors reproduced guideline criteria exactly (verification), while an LLM layer reproduced the same criteria at F1 0.919–0.963 without audit trails.

Conclusions. An asymmetric ≥N-of-M rule improves the specificity of a screening definition and its association with mortality relative to flat any-trigger logic, without improving risk discrimination; component burden relates to mortality in a graded fashion, so ≥3-of-5 functions as a practical rather than a biological threshold. The three-tier taxonomy offers a transparent, auditable, guideline-grounded template for biomarker screening engines.

Key results

Cohort9,245 adults (20–64) · pooled NHANES 2007–2016 · 354 all-cause / 81 CVD deaths

Flat any-trigger (≥1/5)80.4% positive · all-cause HR 1.26 (0.87–1.83) — not significant

Asymmetric ≥3-of-529.2% positive · all-cause HR 1.56 (1.22–1.98) · CVD HR 2.37 (1.39–4.03)

Dose-responseGraded — significant at 4 (HR 2.07) and 5 (HR 2.43) components, not at 3

DiscriminationSimilar across rules (Harrell's C 0.766 vs 0.769)

Detector faithfulnessDeterministic F1 = 1.00 (verification) · LLM layer F1 0.919–0.963

Methods

Deterministic detectors implement guideline thresholds (ATP-III, revised) as pure functions of marker values, emitting machine-readable audit trails (fired branch, trigger strings, confirmatory count). Outcome validation pooled the fasting subsample of five NHANES cycles (2007–2016) linked to the NCHS 2019 Public-Use Linked Mortality File, analysed with survey-weighted Cox models (pooled fasting weights, Taylor-linearized variance, PSU and strata) adjusted for age, sex and race/ethnicity. A cross-sectional cohort (NHANES 2017–2018) assessed rule behaviour and implementation faithfulness versus an LLM screening layer. The full pipeline was implemented in R (survey 4.5; survival 3.8.6) and independently re-run on a separate Linux host with identical results.

Limitations

Metabolic-syndrome components are correlated — the rule encodes a clustered phenotype, not independent-noise filtering.
Metabolic syndrome was the single pattern with a clean asymmetric cluster definition in this dataset.
Cardiovascular events are limited (n=81), widening confidence intervals; follow-up ends in 2019.
Findings are associational, adjusted for a limited covariate set.
Detector faithfulness (F1 = 1.00) is implementation verification by construction, not independent validation of the criteria.

Declarations

EthicsNot applicable — secondary analysis of de-identified public NHANES and linked mortality files.

FundingBootstrapped (Zenlo LLC).

Competing interestsThe author is Founder & CEO of Zenlo LLC.

Data & codeNHANES and NCHS linked mortality files are public; the analysis pipeline is released with the posted preprint.

How to cite Shibakov D. A Three-Tier Marker Framework with Asymmetric Compound Detection Rules for Metabolic Syndrome Screening: Methodology and Validation Against All-Cause and Cardiovascular Mortality in Pooled NHANES (2007–2016). Preprint draft, 2026.

Preprint draft v3.0 · Zenlo Research · not yet posted