Bilateral ABI Fails to Improve 10-Year Mortality Prediction in NHANES
A new exploratory benchmark in NHANES found no meaningful predictive gain from adding bilateral ankle-brachial index to a flexible clinical model for ten-year death from diseases of the heart or cerebrovascular disease. The analysis used public NHANES 1999-2004 examination data linked to the 2019 public-use mortality file, then split the cohort chronologically into development and temporal evaluation sets. In the evaluation sample, the clinical model already performed well, and bilateral ABI did not move the needle on either discrimination or overall probabilistic accuracy.
The finding is clinically important because ABI has long been treated as a compact marker of systemic atherosclerosis and peripheral arterial disease. A low ABI still carries pathobiologic meaning: it reflects arterial narrowing, impaired distal perfusion, and diffuse vascular injury. Yet prognostic plausibility does not automatically translate into incremental value once a model already captures age, comorbidity burden, and other clinical predictors. This study asked exactly that narrower question, and the answer was unsupportive.
A single extra vascular measurement did not rescue a model that was already near its ceiling.
A Chronological Design Built to Test Incremental Information
Investigators rebuilt the public NHANES dataset and restricted the analysis to adults aged 40 years or older with bilateral ABI values between 0.20 and 2.50 and ascertainable 120-month vital status. The corrected cohort included 6,807 participants, with 4,617 in development and 2,190 in temporal evaluation. Across the full analytic sample, there were 409 primary events, including 139 in the evaluation period. The model development strategy relied on 1999-2002 data, while assessment was reserved for 2003-2004, a design that avoided the easy trap of judging a predictor on the same population used to train it.
The primary comparison centered on four approaches: linear clinical logistic regression, flexible clinical logistic regression, flexible clinical plus ABI regression, and histogram gradient boosting plus ABI. Full evaluation weights were used for point estimates, and uncertainty was quantified with 32-replicate Fay balanced repeated replication and 15 design degrees of freedom. That combination of survey weighting and temporal validation matters because NHANES is not a simple convenience sample; its estimates need to reflect the survey architecture if the results are to remain credible.
Flexible Clinical Modeling Already Captured Most of the Signal
In the evaluation cohort, the flexible clinical regression model reached an AUROC of 0.834. Adding bilateral ABI did not change that figure: AUROC remained 0.834 after ABI was appended. The paired ABI increment was -0.0007, with a 95% confidence interval from -0.0047 to 0.0032. For calibration-like performance, the paired Brier difference was 0.0008, with a 95% confidence interval from -0.0003 to 0.0018. Neither result supported the hypothesis that ABI would improve both discrimination and Brier score beyond the clinical benchmark.
The pathology behind ABI makes the null result more interesting, not less. ABI should, in theory, act as a distal readout of generalized atherosclerotic burden and vascular stiffness. In a population-based cohort, however, much of that risk may already be encoded indirectly through age, blood pressure-related variables, smoking-related injury, diabetes, prior vascular disease, and related clinical features. When a flexible model is allowed to absorb nonlinear relationships among these factors, the remaining information contained in ABI may be too small or too redundant to change prediction materially.
Boosting Did Not Outperform the Flexible Clinical Benchmark
The second comparison asked whether a limited nonlinear learner could improve the same chronological evaluation. Histogram gradient boosting plus ABI produced an AUROC of 0.822. Versus the flexible regression model using the same predictor block, the AUROC difference was -0.0117, with a 95% confidence interval from -0.0379 to 0.0144. The Brier difference was -0.000661, with a 95% confidence interval from -0.001322 to -7.53e-08; the upper confidence bound came very close to zero.
Those numbers do not point to a superior machine-learning rescue. They instead suggest that, in this specific benchmark, greater algorithmic complexity did not reveal hidden prognostic structure. In practical terms, a more flexible learner could not extract enough extra signal from the ABI-augmented feature set to beat the already strong clinical model. For a bedside test often viewed as an easy add-on to routine vascular assessment, that is a sobering result.
What the Negative Result Does and Does Not Mean
The investigators were careful not to overstate the conclusion. Bilateral ABI remained associated with vascular disease in the broader clinical sense, and this analysis does not negate that association. It also does not establish equivalence between models or justify treatment decisions based on these predictions. What it does show is narrower: within this corrected, exploratory temporal benchmark, bilateral ABI did not improve prediction of ten-year heart or cerebrovascular mortality beyond a matched flexible clinical model.
That distinction matters because prognostic association and incremental predictive utility are not interchangeable. A biomarker can be biologically valid and still add little once stronger or overlapping predictors are already in the model. The present results fit that pattern. ABI may still help in diagnostic workups for peripheral arterial disease or in identifying systemic atherosclerosis, but this preprint offers no support for using it as a meaningful enhancer of an already competent mortality prediction tool in the NHANES setting.
Why the Benchmark Matters for Cardiovascular Risk Modeling
The study also serves as a reminder that predictive claims need a strict temporal test. Cross-sectional fit, in-sample improvement, or retrospective optimism can make almost any candidate marker look useful. Temporal validation is a harder standard. It asks whether the predictor retains value when applied to a later slice of the population, after model development has already occurred. Here, bilateral ABI did not clear that bar.
For cardiovascular epidemiology, the result reinforces a basic modeling principle: incremental value should be demonstrated against a relevant comparator, for a concrete endpoint, and under a validation design that approximates future use. Without that discipline, a biomarker can appear more useful than it is. In this analysis, the comparator was not a trivial baseline. It was a flexible clinical model already capable of strong performance, which makes the absence of ABI gain more informative than a weaker null result would have been.
The study was not prospectively registered, and the authors note that the analysis used openly available, de-identified data from NHANES and the linked mortality file. A persistent repository deposit is pending, and the linked participant-level records were not uploaded to medRxiv. Those transparency details strengthen reproducibility, but they do not change the substantive message: in this exploratory benchmark, bilateral ABI did not improve ten-year mortality prediction in a way that would support clinical utility claims.