Machine Learning Models Fail to Outperform Benchmark in Predicting Arterial Stiffness Changes

A hybrid modeling approach combining machine learning and penalized regression failed to outperform a simple mean-only benchmark when predicting five-year changes in carotid-femoral pulse wave velocity, according to an exploratory prospective study published in BMC Medical Informatics and Decision Making. Investigating the longitudinal progression of vascular aging in a population-based cohort, the research team evaluated whether selected baseline clinical and lifestyle variables could reliably predict arterial stiffening over a five-year follow-up period, finding that neither primary analytical framework improved predictive accuracy beyond basic mean estimates.

Arterial stiffness, measured clinically through carotid-femoral pulse wave velocity (cf-PWV), serves as an established and validated biomarker of vascular aging and cardiovascular risk. Although cf-PWV is widely recognized for assessing vascular health, the extent to which baseline clinical and behavioral factors can forecast longitudinal changes in this metric has remained poorly understood.

The study analyzed data from the Influence of Different Risk Factors in Vascular Accelerated Aging cohort, also known as the Estudio de Envejecimiento Vascular Acelerado (EVA study). This prospective cohort investigation enrolled adult participants who were completely free from cardiovascular disease at baseline. The trial protocol was retrospectively registered under ClinicalTrials.gov identifier NCT02623894 on December 8, 2015, providing a prospective framework to examine how risk factors influence vascular aging trajectories over time.

To capture potential nonlinear interactions and establish predictor importance, the researchers employed an Extreme Gradient Boosting (XGBoost) algorithm as their initial analytical stage. Model interpretability was evaluated using SHapley Additive exPlanations (SHAP) analysis, which allowed the team to quantify the relative predictive contributions of individual features. Following this machine learning step, Elastic Net regression was applied to derive a parsimonious and clinically interpretable predictive equation. Model tuning was systematically conducted using cross-validation within the training set, and overall predictive performance was subsequently verified on a separate, held-out test set.

The preliminary XGBoost analysis identified six primary baseline variables as key predictors of five-year cf-PWV change: smoking duration in years, adherence to the Mediterranean diet, high-density lipoprotein (HDL) cholesterol, the use of antidiabetic drugs, baseline age, and weekly sitting time measured in hours per week. When ranking features by mean absolute SHAP values, the largest relative contributions were driven by smoking duration, chronological baseline age, and weekly sitting duration.

The hierarchical Elastic Net regression model subsequently retained a specific subset of main effects and interaction terms. Chronological age at baseline was retained as a positive main effect with a standardized beta coefficient of 0.31, while weekly sitting time contributed a positive main effect beta of 0.02. Conversely, the model retained negative coefficients for HDL cholesterol (beta = −0.13), use of antidiabetic drugs (beta = −0.11), and adherence to the Mediterranean diet (beta = −0.008). In addition to these primary variables, the model incorporated an interaction term between baseline age and weekly sitting time, which was assigned a negative beta coefficient of −0.24.

When evaluated against the held-out test data, the hierarchical Elastic Net model demonstrated limited predictive performance. The derived equation yielded a mean absolute error (MAE) of 1.27 meters per second and a root mean square error (RMSE) of 1.70 meters per second. These results showed no predictive advantage when compared with a simple mean-only benchmark, which recorded an identical MAE of 1.27 meters per second and a marginally superior RMSE of 1.68 meters per second.

Because neither the XGBoost approach nor the penalized Elastic Net regression model improved upon the reference benchmark, the study authors concluded that the derived mathematical equation cannot serve as a definitive predictive tool. Instead, the authors noted that the equation should be treated as an exploratory, hypothesis-generating research model that requires further methodological refinement and independent external validation in separate cohorts.

Subject of Research: Predicting five-year changes in carotid-femoral pulse wave velocity using machine learning and penalized regression

Article Title: Predictive contributions to five-year change in carotid-femoral pulse wave velocity using machine learning and penalized regression

Article References: Patino-Alonso, C., Gómez-Sánchez, M., Rodríguez-Sánchez, E., García-Ortiz, L., Gómez-Sánchez, L., & Gómez-Marcos, M. A. (2026). Predictive contributions to five-year change in carotid-femoral pulse wave velocity using machine learning and penalized regression. BMC Medical Informatics and Decision Making. https://doi.org/10.1186/s12911-026-03888-1

Image Credits: AI Generated

DOI: 10.1186/s12911-026-03888-1

Keywords: carotid-femoral pulse wave velocity, arterial stiffness, vascular aging, cardiovascular risk, XGBoost, Elastic Net regression, SHAP analysis, EVA study, smoking duration, Mediterranean diet, HDL cholesterol, antidiabetic drugs, sitting time, machine learning, penalized regression, prospective cohort, cross-validation