This review article examines the application of deep learning in electroencephalogram (EEG) classification tasks, aiming to address critical questions such as which tasks have been explored, what input formulations have been used, and whether specific deep learning network structures are suitable for specific tasks. The authors conducted a systematic literature review using Web of Science and PubMed databases, identifying 90 studies. These studies were analyzed based on task type, EEG preprocessing methods, input type, and deep learning architecture. The main findings include: 1. Convolutional neural networks (CNNs), recurrent neural networks (RNNs), and deep belief networks (DBNs) outperform stacked auto-encoders and multi-layer perceptron neural networks in classification accuracy. 2. EEG classification tasks were categorized into five general groups: emotion recognition, motor imagery, mental workload, seizure detection, and event-related potential detection. 3. Preprocessing methods for EEG data included artifact removal strategies such as manual, automatic, and no cleaning, with frequency domain filters used to limit the bandwidth of the EEG. 4. Input formulations for training deep networks included calculated features, images, and signal values, with CNNs and DBNs being the most prevalent. 5. Deep learning architecture trends showed that CNNs were the most common, followed by DBNs, hybrid architectures, RNNs, and MLPNNs. 6. Activation functions were predominantly used in convolutional layers, with ReLU being the most popular. 7. Task-specific deep learning trends indicated that emotion recognition, motor imagery, and sleep stage scoring tasks did not show a clear consensus on deep learning algorithms, while seizure detection studies favored CNNs or RNNs, and ERP studies preferred CNNs. The review provides recommendations for design choices based on task type, including specific input formulations and architecture parameters. It also discusses hybrid architecture types and their performance. The overall goal is to guide future research and applications of deep learning in EEG classification.This review article examines the application of deep learning in electroencephalogram (EEG) classification tasks, aiming to address critical questions such as which tasks have been explored, what input formulations have been used, and whether specific deep learning network structures are suitable for specific tasks. The authors conducted a systematic literature review using Web of Science and PubMed databases, identifying 90 studies. These studies were analyzed based on task type, EEG preprocessing methods, input type, and deep learning architecture. The main findings include: 1. Convolutional neural networks (CNNs), recurrent neural networks (RNNs), and deep belief networks (DBNs) outperform stacked auto-encoders and multi-layer perceptron neural networks in classification accuracy. 2. EEG classification tasks were categorized into five general groups: emotion recognition, motor imagery, mental workload, seizure detection, and event-related potential detection. 3. Preprocessing methods for EEG data included artifact removal strategies such as manual, automatic, and no cleaning, with frequency domain filters used to limit the bandwidth of the EEG. 4. Input formulations for training deep networks included calculated features, images, and signal values, with CNNs and DBNs being the most prevalent. 5. Deep learning architecture trends showed that CNNs were the most common, followed by DBNs, hybrid architectures, RNNs, and MLPNNs. 6. Activation functions were predominantly used in convolutional layers, with ReLU being the most popular. 7. Task-specific deep learning trends indicated that emotion recognition, motor imagery, and sleep stage scoring tasks did not show a clear consensus on deep learning algorithms, while seizure detection studies favored CNNs or RNNs, and ERP studies preferred CNNs. The review provides recommendations for design choices based on task type, including specific input formulations and architecture parameters. It also discusses hybrid architecture types and their performance. The overall goal is to guide future research and applications of deep learning in EEG classification.