Background N-terminal pro-B-type natriuretic peptides (NT-proBNP) are important in the assessment of suspected heart failure (HF). However, NT-proBNP concentrations are elevated in atrial fibrillation (AF), creating diagnostic uncertainty. The aim of this study was to assess the diagnostic accuracy of NT-proBNP for HF in people with AF, overall and by age, sex and BMI. Methods and findings Retrospective study of all patients with a NT-proBNP test in their primary care electronic health record among English GP practices provided through the Clinical Practice Research Datalink (2004–2018) and linked to secondary care data. The accuracy of NT-proBNP for diagnosing HF within six months was assessed for people with and without AF at thresholds of 125, 400, 660 and 2,000 pg/mL, including by age, sex and BMI. Among 155,347 people who had an NT-proBNP test organized in primary care (median age 61 years), 17,403 (11.2%) had pre-existing AF. Of the 155,347 people included, 14,585 (9.4%) were subsequently diagnosed with HF, including 4,168/17,403 (23.9%) people with AF (median NT-proBNP = 1,852 pg/mL, interquartile range (IQR) [974, 3,459] pg/mL) and 10,417/137,944 (7.6%) without AF (1,110 pg/mL, IQR [434, 3,108] pg/mL). NT-proBNP discriminated better overall among people without AF (AUC = 0.877 (95% confidence interval (CI) [0.873, 0.881]) than with AF (AUC = 0.743 (95% CI [0.735, 0.751]). Among people with AF, NT-proBNP sensitivity and specificity at a 125 pg/mL threshold was 98.8% (95% CI [98.5%, 99.1]) and 13.2% (95% CI [12.6%, 13.7]) and at 400 pg/mL 93.2% (95% CI [92.4, 93.9]) and 35.5% (95% CI [34.7, 36.3]). Among people without AF the corresponding results were 92.9% (95% CI [92.4, 93.4]) and 53.8% (95% CI [53.6, 54.1]) at 125 pg/mL and 77.1% (95% CI [76.3, 77.9]) and 84.9% (95% CI [84.7, 85.1]) at 400 pg/mL. NT-proBNP discriminated less well among people with AF aged ≥65 years compared to <65years (e.g., AUC in people aged 65–75 years was 0.725, 95% CI [0.712, 0.739]). Increasing the threshold for a positive test among people with AF from 125 pg/mL to 660 pg/mL would reduce the number of false positive results by 26.0%, whilst retaining a negative predictive value of 91.5 (95% CI [90.8, 92.1]), albeit with a 10.6% increase in the proportion of those tested with AF having a missed or delayed HF diagnosis. The main limitation of the study is that it relies on routinely collected primary care data and people with an NT-proBNP result <400 pg/ mL may not have been referred for further assessment, impacting upon the diagnostic accuracy below this threshold. Conclusions NT-proBNP discriminates more accurately for HF among people without AF than with AF. A higher referral threshold could be considered in AF to account for higher median NT-proBNP levels but this would also increase missed HF diagnoses.