The　analysis　of　electrocardiogram　(ECG)　signals　holds　the　potential　to　predict　arrhythmia　and　prevent　cardiovas-cular　disease.　However,　ECG　data　possess　high　dimensionality,　high　noise,　and　exhibit　time　sequence　properties.　Traditional　machine　learning　algorithms　often　eliminate　many　highly　correlated　features　during　dimensionality　reduction,　thereby　disregarding　the　temporal　aspect　of　the　data　and　resulting　in　subpar　detection　performance.　In　this　paper,　we　propose　a　novel　arrhythmia　detection　method　from　ECG　signals,　named　WTDRL,　which　combines　wavelet　transform　and　deep　reinforcement　learning.　Specifically,　the　proposed　method　employs　a　discrete　wavelet　transform　to　decompose　the　signal　into　sub-signals　characterized　by　distinct　frequency　ranges.　By　subsequently　applying　suitable　filtering　and　thresholding　techniques　to　each　sub-signal,　the　presence　of　noise　can　be　effectively　mitigated.　Furthermore,　WTDRL　leverages　a　reinforcement　learning　model　with　the　deep　deterministic　strategy　algorithm　and　long　short-term　memory　(LSTM)　to　determine　the　optimal　strategy　for　predicting　the　heartbeat　type　of　subjects　and　effectively　detecting　arrhythmia.　Experimental　results　on　a　public　ECG　dataset　demonstrate　that　the　proposed　method　surpasses　traditional　approaches　and　some　state-of-the-art　methods　of　deep　learning　in　terms　of　arrhythmia　detection　performance.　©　2023　IEEE.