Development and external validation of a clinical model to estimate symptom-limited cardiopulmonary exercise testing parameters for prognostication in heart failure

Background Cardiopulmonary exercise testing (CPET) yields strong prognostic predictors in heart failure (HF)—oxygen uptake at the anaerobic threshold (ATVO₂), peak oxygen uptake (peakVO₂), and the percentage of predicted peakVO₂ (peakVO₂%pred)—but these are often unavailable clinically due to disease severity, resource constraints, or lack of testing, creating a risk stratification gap for many patients. Objective To develop a clinical model using routinely available variables to estimate key CPET parameters and to evaluate whether these estimated values retain prognostic value for mortality in an external population. Methods This study included 209 patients with NYHA class III–IV HF who underwent symptom-limited CPET. Echocardiography was performed at baseline and 6-month follow-up. The primary outcome was a 6-level hierarchical composite endpoint (all-cause death/heart transplantation, HF hospitalization, echocardiographic deterioration, persistent abnormality, improvement without normalization, and maintained/restored normal echocardiography). LASSO regression selected predictors from routinely available variables (age, sex, BMI, iron deficiency, uric acid, log-hsCRP, hypertension, diabetes) to estimate ATVO₂, peakVO₂, and peakVO₂%pred. The models were externally validated in 1,931 NHANES 2017–2018 participants. Cox models assessed associations between estimated CPET values and all-cause mortality. RCS regression compared linear trends across cohorts, and subgroup analysis evaluated prognostic performance in different populations. Results Over a median 458-day follow-up, in the models derived from COX-LASSO regression, ATVO₂, peakVO₂, and peakVO₂%pred were all significant predictors of the outcome (all p<0.001). LASSO selected variable subsets for each parameter: ATVO₂ and peakVO₂ were predicted by iron deficiency, age, uric acid, log-hsCRP, and diabetes; peakVO₂%pred by iron deficiency, age, sex, uric acid, and hypertension. In the NHANES cohort, estimated ATVO₂ (HR: 0.107; 95%CI: 0.031–0.363; p<0.001) and estimated peakVO₂ (HR: 0.453; 95%CI: 0.252–0.814; p<0.05) significantly predicted all-cause mortality, whereas estimated peakVO₂%pred did not. RCS curves demonstrated similar linear trends for ATVO₂ and peakVO₂ between cohorts. In NHANES subgroup analysis, estimated ATVO₂ and peakVO₂ did not predict mortality in those >52 years, nor did peakVO₂ in males. However, estimated peakVO₂%pred remained significant in females and males. Conclusions Routinely available variables estimate ATVO₂ and peakVO₂, both significantly predicting general population mortality. Estimated peakVO₂%pred showed limited generalizability but remained significant in females and males on subgroup analysis. These findings affirm the prognostic value of ATVO₂ and peakVO₂ and support broader CPET use. Larger studies are needed to refine estimation models when direct CPET is unavailable.