BACKGROUND Patients with atrial fibrillation (AF) have a higher risk of cognitive impairment (CI). However, complexity of CI diagnosis and lack of simple screening approaches limited early screening and intervention of CI in AF patients. OBJECTIVE Our study aimed to develop deep learning models based on fundus photographs for easy screening of CI in AF patients. METHODS From May 2021 to April 2023, patients who completed fundus examination and cognitive function evaluation in the Chinese Atrial Fibrillation Registry Study were included. The training and validation sets were randomly split at an 8:2 ratio. Participants from the Beijing Eye Study served as the external validation set. Different deep learning models were trained, and their CI detection ability was validated. RESULTS A total of 899 patients in the Chinese Atrial Fibrillation Registry Study were included. In the validation set, the vision-ensemble model based on fundus images alone had an area under the receiver-operating characteristic curve (AUROC) of 0.855 (95% confidence interval 0.816-0.894) for CI screening. The multimodal model (AUROC 0.861, 95% confidence interval 0.823-0.898), based on fundus photographs and 4 clinical variables, performed comparably to the vision-ensemble model. The AUROC of the vision-ensemble model for CI screening achieved 0.773 (95% confidence interval 0.709-0.837) in the external test set. In the saliency map, the vision-ensemble model focused on areas around retinal vessels and the optic disc. CONCLUSION A vision-ensemble model based on fundus images might be practical for preliminary screening of CI in AF patients.