Electroencephalogram (EEG)-based epilepsy detection has always been a challenging problem for both doctors and researchers in recent years because of its complexity. Although many works have been conducted, some of which reached significantly high detection accuracy using datasets (for example, CHB-MIT scalp EEG database) available on the Internet, limitations still exist, mainly coming from the inability to find common sets of features across patients. Furthermore, determining which features are effective and how many effective features should be included in the detection is another challenge for researchers. In this paper, the authors propose an approach to investigate which features are effective and which number of effective features are optimal to achieve the best results using shallow learning. In addition, to overcome overfitting in shallow learning when multiple features are exploited, this study also conducts a combination of Deep Learning (DL) and large-scale features of many domains like time domain, frequency domain, time-frequency domain, and wavelet domain to classify seizure state from interictal state. Firstly, Fast Fourier Transform (FFT), Short-Time Fourier Transform (STFT), and Multilevel Discrete Wavelet Transform (MDWT) are used to extract features. Secondly, the tree-based feature selection is applied to choose the ten most relevant features and then combined with the Random Forest (RF), resulting in an accuracy above 99,1%. Finally, a Convolutional Neural Network (CNN) is implemented to automatically classify feature images of ictal and interictal states. The results of this experiment are above 95% for the patient who has the highest ictal interval. The number of ictal images is a little for other patients, and their accuracies are lower. This is a promising sign that if the database is larger, multi-domain features can provide high accuracy for seizure detection. The proposed method of the paper shows that the use of EEG signals in many domains is essential to increase the accuracy of epilepsy detection in both traditional machine learning and deep learning. Feeding plenty of features into DL effectively prevents overfitting in shallow learning and detects seizures across patients, thus improving training time.