An apparatus for treating a person having cardiac arrest includes a processor and a lead in communication with the processor, the lead being configured to receive a segment of electrocardiogram (ECG) data from the person while undergoing cardiopulmonary resuscitation (CPR). The apparatus also includes a non-transitory computer-readable medium having: a first Convolutional Neural Network-based Encoder-Decoder (CNNED) structure to reduce artifacts due to the CPR; a second CNNED structure cascaded with the first CNNED structure to further reduce artifacts; and instructions to determine a first arrhythmia classification using the first CNNED structure, determine a second arrhythmia classification using the second CNNED structure in response to the first arrhythmia classification indicating a shockable arrhythmia event; and transmit a signal to apply a defibrillation shock to the person in response to the first arrhythmia classification and the second arrhythmia classification indicating a shockable arrhythmia of the person.