Background: Early detection of atrial fibrillation (AF) and subsequent initiation of anticoagulation and rhythm control therapy markedly reduce stroke, cardiovascular death, and heart failure, but unselected ECG screening for AF is time- and resource-intensive. Combining biomarkers reflecting different biological processes may identify patients at high risk of AF, enabling targeted ECG screening. Methods: We systematically reviewed literature and patent information to define candidate biomarkers for AF detection. The top 12 biomarkers identified through this process were quantified on a high-precision, high-throughput platform in 1485 consecutive patients at risk for AF presenting acutely to hospital (median age 69 years [Q1, Q3 60,78]; 60% male). Patients had either known AF (45%) or AF ruled out by 7-day ECG-monitoring. A model simultaneously considering 7 key clinical characteristics and all biomarkers was developed in a randomly sampled discovery cohort (n=933) and validated in the remaining patients (n=552). Neural networks were also applied. Findings: Using backward elimination, a model using age, sex, body mass index (BMI), BMP10, ANG2, and FGF23 discriminated between patients with and without AF with an AUC of 0·743 [95% confidence interval (CI) 0·712-0·775]. The biomarkers represent distinct pathways relevant for atrial cardiomyopathy and AF, namely hypertrophy and fibrosis (FGF23), endothelial dysfunction (ANG2), and atrial oxidative stress and the genomic predisposition to AF (BMP10). The SHAP procedure for neural networks identified the same variables as the regression. The validation yielded an AUC of 0·719 (95%CI 0·677, 0·762), corroborated using deep neural networks (AUC 0·784 [95%CI 0·745, 0·822]). Interpretation: The combination of three simple characteristics (age, sex, BMI) and three biomarkers (BMP10, ANG2, and FGF23) robustly identifies patients with AF. Such an approach enables targeted screening for AF and provides a platform to develop personalised prevention and treatment in patients with AF. Funding Statement: This work was partially supported by the European Commission (grant agreements no. 633196 [CATCH ME]) to PKi, LF, BC, SH, SK, LM, MFS, US, and no. 116074 [BigData@Heart EU IMI] to PKi, British Heart Foundation (FS/13/43/30324 and (AA/18/2/34218) to PKi and LF), German Centre for Cardiovascular Research supported by the German Ministry of Education and Research (DZHK, via a grant to AFNET to PKi), and Leducq Foundation (14CVD01) to PKi. Declaration of Interests: LF has received institutional research grants and non-financial support from European Union, British Heart Foundation, Medical Research Council (UK), and DFG and Gilead. PKi has received additional support for research from the European Union, British Heart Foundation, Leducq Foundation, Medical Research Council (UK), and German Centre for Heart Research, from several drug and device companies active in atrial fibrillation, honoraria from several such companies. PKi and LF are listed as inventors on two patents held by University of Birmingham (Atrial Fibrillation Therapy WO 2015140571, Markers for Atrial Fibrillation WO 2016012783). PKa is an employee of Roche Diagnostics GmbH. AZ is an employee of Roche Diagnostics Intl. All other authors have reported no relationships relevant to the contents of this paper to disclose. Ethics Approval Statement: This study complied with the Declaration of Helsinki, was approved by the National Research Ethics Service Committee (IRAS ID 97753), and was sponsored by the University of Birmingham.