*3.3. Datasets*

Figure 7 presents the usage percentage of datasets used in emotion recognition. DEEP and SEED are publicly available databases, and are the most frequently used (49% and 23% of applications, respectively). Sometimes, other studies used self-generated datasets (23%), which are typically not freely accessible. The MAHNOB-HCI and RCLS public datasets appeared in our research sample, with a participation of 3% each.

**Figure 7.** EEG datasets for emotion recognition.

Systems that use public databases o ffer some comparability, but contrast is limited even if the same characteristics are handled. Still, such public databases could eventually lead to findings if objective comparisons are performed.

#### *3.4. Feature Extraction*

Most systems use feature extraction methods in the time, frequency, time-frequency, or space domains. A small percentage of works evaluate the functional connectivity (or di fferences) in the observed activity between brain regions when emotions are provoked. Features with non-redundant information combined from di fferent domains yield better classification results. However, it is still unclear if features work better alone or in combination with each other, or which type of features are more relevant for emotion recognition.

In our review, we found that researchers addressed these issues through the development of feature extraction algorithms that outperform the classic frequency bands and extract as much information as possible from brain signals. We believe that further developments should be connected to a comprehensive understanding of the brain's neurophysiology.

Figure 8 presents the domains of the used features. Frequency domain features are the most frequently used, and appear nearly twice as often as time domain or time-frequency domain features. Asymmetry characteristics between electrode pairs (by each hemisphere) are increasingly being used—likewise, electrodes' location data in di fferent brain sections. Additionally, raw data (without features) is used as inputs for deep learning classifiers.

**Figure 8.** Domain of used features.

Figure 9 shows the usage percentage of various algorithms for feature extraction computed in the 31 papers shown in Table 8. We found that FFT, SFFT, and DFT are the most commonly used tools for characteristic extraction in the frequency domain (27.9%). AR is used less frequently to estimate the spectrum (4.7%). WT and DWT appear in 23.3% of the systems in our sample. These algorithms are applied to obtain features in the time-frequency domain. Likewise, data from channel or electrode specific locations are less frequent (4.7%). Researchers also use statistics and computed parameters in the time domain (9.3%), normalized mutual information NMI (2.3%), ERS (2.3%), and ERD (2.3%).

**Figure 9.** Percentage of the use of algorithms for feature extraction from Table 8.

We observed an increasing presence of algorithms embedded in neural networks like RBN, DBN, TensorFlow functions, and LSTM (4.7%) that are used to extract signal features automatically from raw data. This approach yields a good enough classifier performance, probably because it preserves information and avoids the risk of removing essential emotion-related signal features.

#### *3.5. Feature Selection*

It is worth noting that 61.3% of the systems presented in Table 8 do not use a feature selection method. Table 9 lists the systems that utilized feature selection algorithms. Interestingly, virtually every system uses a di fferent algorithm except for the methods minimum redundancy maximum relevance (mRMR) and recursive feature elimination, which are utilized for two di fferent schemes.


**Table 9.** Systems in Table 8 using feature selection algorithms.
