Personalized evaluation of NT-proBNP and clinical characteristics in suspected heart failure: the PRECISE-HF study

N-terminal pro-B-type natriuretic peptide (NT-proBNP) is recommended to guide echocardiography referral in community-based patients with suspected heart failure (HF), but diagnostic performance is influenced by individual characteristics including age, ethnicity, kidney function, cardiac rhythm, and body mass index (BMI). An individualised probabilistic approach may have superior diagnostic performance to guideline-recommended universal thresholds. This study aimed to develop and validate the PRECISE-HF model to provide probabilistic rule-out and rule-in thresholds based on NT-proBNP and relevant clinical characteristics.

Cohort study performed from January 1, 2018 to March 31, 2025, using primary care data from the Clinical Practice Research Datalink (CPRD), broadly representative of the UK population. Patients were included if they had an NT-proBNP measured due to clinical suspicion of HF with no known history of HF. HF diagnosis was recorded as a primary care diagnosis or HF hospitalisation within 12 months of NT-proBNP measurement. Using XGBoost machine learning, the probability of HF was modelled incorporating NT-proBNP alongside age, sex, ethnicity, estimated glomerular filtration rate, BMI, systolic blood pressure, anaemia, loop diuretic prescription and history of atrial fibrillation, diabetes, myocardial infarction and chronic obstructive pulmonary disease. Rule-out and rule-in thresholds were identified based on ≥90% sensitivity and specificity, respectively, and compared to current European Society of Cardiology (ESC) recommendations. External validation was performed in the REVOLUTION-HF Swedish cohort.

Overall, 535 583 patients were included, randomly split into derivation (n=374 909) and internal validation (n=160 674) cohorts. A HF diagnosis was recorded in 10% of the validation cohort. The PRECISE-HF model demonstrated excellent discrimination (area under the receiver operating characteristic curve [AUROC] 0.896) and calibration (Brier score 0.061). The rule-out threshold had a sensitivity of 90.1% and a negative predictive value of 98.5%, ruling out 64.8% of patients. The rule-in threshold ruled in 16.1% with a specificity of 90.0% and a positive predictive value of 44.1%. Compared to the ESC thresholds, PRECISE-HF13 ruled out 144 additional patients per 1000 at the expense of three missed diagnoses, with 73 fewer false positives per 1000 at rule-in. In external validation, the model retained good discrimination (AUROC 0.757) and calibration (Brier score 0.163).

The PRECISE-HF model, integrating multiple clinical characteristics with NT-proBNP, demonstrated superior diagnostic performance compared with universal rule-out and age-adjusted rule-in thresholds in community-based patients with suspected HF. This individualised approach may enable earlier confirmatory testing and initiation of disease-modifying treatment in those most likely to have HF.