Sudden cardiac arrest (SCA) is a medical emergency that poses the risk of death to the patient. For prevention, a robust system is intensively on ongoing research to predict whether a person may predict sudden cardiac arrest in the future, long before the incident. The current system development method is learning from the patient's ECG recordings. A combination method between the Wavelet Packet Transform and the Convolutional Neural Network classification model was proposed to obtain hidden patterns from the patient's ECG recordings. The SCA dataset was obtained from the MIT-BIH SCA Holter DB with 2 ECG machines and the Normal Dataset from the MIT-BIH Normal Sinus Rhyme. With cross-validation evaluation with kfold=10, by comparing one segment 1 minute from 30 minutes before onset, Convolutional Neural Network (CNN) performed with an accuracy of 95.89%, precision of 96.75%, recall 95.83%, and F1-Score 95.75%, with two-level Meyer Wavelet Packet Transformation.