Electrocardiograph　(ECG)　signals　are　an　important　source　of　data　on　human　heart　health　and　are　widely　used　to　detect　different　types　of　arrhythmias.　With　the　development　of　deep　learning,　end-to-end　ECG　classification　models　based　on　neural　networks　have　emerged,　but　most　of　the　existing　models　are　only　for　small　datasets,　and　there　is　a　problem　of　poor　generalization　ability.　In　order　to　solve　the　above　two　problems,　we　present　a　novel　deep　learning　architecture,　DRNet,　which　integrates　DenseNet　with　bidirectional　long　short-term　memory　(BiLSTM)　to　form　the　Dense　(D)　Recurrent　(R)　Network　with　BiLSTM　serving　as　the　recurrent　structure.　The　DenseNet　layer　is　used　to　extract　the　deep　features　of　the　ECG　signal;　and　the　BiLSTM　layer　can　aggregate　the　extracted　features　in　time,　which　is　more　sensitive　to　the　timing　status　of　the　ECG.　These　deep　features　enrich　the　temporal　background　and　serve　as　the　foundation　for　subsequent　classification　tasks.　In　the　classification　stage,　the　model　categorizes　ECG　signals　into　one　of　four　types:　＂normal,＂　＂AF＂　(atrial　fibrillation),　＂other　rhythms,＂　and　＂noisy.＂　In　addition,　in　order　to　solve　the　problem　of　misclassification　caused　by　imbalanced　ECG　signal　samples,　an　improved　cross-entropy　loss　function,　a　multiclass　concentration　loss　function,　is　proposed　in　this　paper,　and　the　partial　derivative　solution　process　is　given.　Finally,　in　the　prediction　of　ECG　signals,　the　existing　simple　voting　is　improved　to　a　weighted　voting　method.　The　comparison　experiment　with　the　ordinary　dense　neural　network　on　the　published　larger　ECG　dataset　proves　that　this　method　not　only　simplifies　feature　extraction　but　also　greatly　enhances　the　generalization　ability　of　the　network.　©　The　Author(s),　under　exclusive　licence　to　Springer-Verlag　London　Ltd.,　part　of　Springer　Nature　2024.