Atrial　fibrillation　(AF)　is　one　of　the　most　common　kind　of　arrhythmia　and　its　incidence　and　prevalence　are　increasing　worldwide,　with　significant　public　health　implications.　As　a　result,　increasing　attention　has　been　paid　to　predicting　AF　in　order　to　prevent　further　progression　of　the　disease.　In　this　paper,　we　introduce　a　novel　deep　learning　classification　method　that　combines　DenseNet　and　Transformer,　namely　DenseTransformer,　to　implement　classification　for　AF　prediction　using　short　electrocardiogram　(ECG)　signals.　DenseNet　architecture　is　employed　to　extract　local　features　and　encourage　feature　reuse.　Our　model　incorporates　a　novel　bidirectional　transformation　mechanism,　capturing　global　features　and　temporal　dependency　from　both　preceding　and　following　context.　We　develop　and　evaluate　the　proposed　method　based　on　LTAFDB　and　NSRDB,　which　are　available　from　the　PhysioNet.　The　proposed　method　achieved　97.2%　accuracy　in　predicting　AF.　The　results　indicate　that　the　proposed　method　can　accurately　predict　AF　from　short　ECG　signals.　It　can　be　foreseen　that　the　proposed　model　shows　great　potential　for　use　in　clinical　auxiliary　diagnosis　based　on　short　ECG　signals.　©　2025　IEEE.