Introduction: Atrial fibrillation (AF) is the most common arrhythmia affecting millions of individuals worldwide and usually causes serious complications such as stroke. The paroxysmal nature of AF in its early stage makes it especially difficult to diagnose with ambulatory electrocardiogram (ECG) monitors. The increasing availability of wearable photoplethysmographic (PPG) devices appears as a promising solution for AF screening in large populations. This study evaluates the performance of a PPG peak detector and a recurrent neural network to detect AF from PPG signals collected with a wrist-worn device. Methods: Simultaneous PPG and 12-lead ECG recording were analyzed in 21 patients referred for catheter ablation of cardiac arrhythmias. The ECG signals were automatically labelled as AF and no-AF episodes (sinus rhythm (SR), supra- and ventricular arrhythmias) by a system composed of a QRS detector (to extract RR intervals) and a classifier, which was validated against expert’s annotations. After aligning the PPG and ECG signals, the probability of 30s windows of PPG data to be AF was estimated using a stepwise algorithm composed of: a pulse detector extracting interbeat intervals; a quality index estimator discarding non-consistent pulses; a neural network with a gated recurrent unit and a logistic layer detecting AF. The neural network was trained on ECG-based RR intervals from the PhysioNet Long-Term AF Database. Performance metrics were computed by comparing PPG-based predictions to the ECG-based reference. Results: 1039 non-overlapping windows (173 AF, 865 no-AF) of PPG signals were analyzed. The algorithm achieved an accuracy of 93% when discriminating AF from no-AF episodes. The sensitivity and specificity were 89% and 94%. Conclusion: Our approach identifies with high accuracy AF episodes from PPG-based sequences containing SR as well as various arrhythmia types. Our algorithm, recently embedded in wearable device, paves the way for the screening of AF with limited upstream interventions.