Automatic　arrhythmia　detection　in　electrocardiogram　(ECG)　aims　to　enable　computer　to　recognize　different　types　of　arrhythmia　to　assist　doctors　in　diagnosis.　In　medical　conclusion,　there　is　lead　specificity　in　different　types　of　arrhythmias,　and　doctors　mainly　identify　different　arrhythmias　according　to　these　specific　manifestations　in　clinical　diagnosis.　However,　most　of　the　existing　methods　focus　on　the　temporal　dimension　of　ECG　signals,　ignoring　the　dependence　of　ECG　signal　leads.　This　work　aims　to　develop　a　general　method　for　multi-lead　ECG　arrhythmia　recognization　through　exploring　lead　feature　extraction　and　fusion.　For　the　first　time,　we　propose　a　novel　Lead-aware　hierarchical　convolutional　neural　network　(LAH-CNN)　that　capture　time　and　lead　features　hierarchically　with　a　lead-leval　attention　module　mining　the　interdependence　between　lead　feature　mappings,　and　then　obtain　Ngram　features　of　leads　for　final　arrhythmia　classification.　Our　proposed　approach　achieved　F1　score　of　78.86%,　99.2%　and　99.1%　on　12-lead　databases　CPCS,　INCART　and　2-lead　database　MIT-BIH　Respectively.　Our　experiments　show　that　N-gram　lead-leval　features　are　an　important　factor　affecting　prediction　performance.　Our　proposed　method　is　competitive　and　achieves　good　robustness　for　arrhythmias　recognization.