EEG Emotion Classification

• EEG emotion classification is the process of decoding affective states from multi-channel scalp recordings by leveraging spectral, temporal, and spatial signal features.
• The approach includes robust signal acquisition, preprocessing (filtering, ICA, normalization), and feature extraction (band power, DE, connectivity) to ensure reliable emotion mapping.
• Advanced deep learning models, such as 3D CNNs, LSTM variants, and transformer-based architectures, improve classification accuracy and interpretability in diverse experimental setups.

Electroencephalography (EEG) emotion classification refers to the computational task of inferring discrete or continuous affective states from multi-channel scalp EEG signals. This field sits at the intersection of affective computing, neuroscience, signal processing, and machine learning, and underpins research and applications in brain–computer interfaces (BCI), human–computer interaction, and behavioral monitoring. EEG emotion classifiers leverage spectral, temporal, and spatial patterns in brain rhythms to decode internal affective states, aiming for robust, generalizable, and interpretable mappings from raw neural activity to emotion labels or dimensions.

1. EEG Signal Acquisition, Preprocessing, and Experiment Design

EEG-based emotion classification systems require both hardware for reliable neural signal acquisition and carefully structured preprocessing to extract informative neural correlates of affective processing.

Hardware Configurations

• Channel layouts: Systems have ranged from low-density (4–8 electrodes) to high-density (32–128+). For example, an 8-channel ADS1299-based setup using P3, Pz, P4, O1, Oz, O2, T5, T6 demonstrates robust four-class emotion recognition in a portable, low-cost setting (Yutian et al., 2024). High-density nets (e.g., 128 channels) provide finer spatial coverage but increase system and computational complexity (Asif et al., 2022).
• Front-ends and platforms: Recent work employs ADCs such as the ADS1299 (8-ch, 24-bit, 5 μV LSB) interfaced with microcontrollers (e.g., ESP32), enabling wireless data streaming via UDP with submillisecond latency (Yutian et al., 2024).
• Consumer devices: Studies increasingly leverage consumer-grade headsets such as Emotiv EPOC (14 channels (Vasquez et al., 26 Dec 2025)), Muse (4 channels (Joshi et al., 2022)), or OpenBCI (8 channels (Lakhan et al., 2018)), which, with modern algorithms, can exceed 90% accuracy in three-way emotion tasks.

Preprocessing Pipelines

• Filtering: Powerline (50/60 Hz) and broadband artifact filtering are standard. Band-pass filters (e.g., Butterworth 5–18 Hz (Yutian et al., 2024) or 0.5–45 Hz (Purwar et al., 31 Jan 2026)) are applied to focus on theta, alpha, and beta bands relevant to affective EEG research.
• Artifact removal: Independent Component Analysis (ICA) is widely used to extract and discard ocular and muscular interferences (Yutian et al., 2024, Asif et al., 2022), sometimes automated via EEGLab pipelines (Shen et al., 2024).
• Epoch segmentation and normalization: Data are segmented around annotated stimuli or subject responses (e.g., emotional event clicks (Asif et al., 2022)), then normalized per channel, often using z-score or amplitude scaling for cross-subject comparability.
• Time–frequency transformation: Short-time Fourier Transform (STFT) or wavelet decompositions are used to construct tensor representations (e.g., channel × time × frequency), essential for architectures that ingest multi-dimensional signal features (Yutian et al., 2024, Asif et al., 2022).

Experimental Designs

• Emotional stimuli: Standardized video clips, music, or VR avatars displaying emotions are used to elicit target affective states, with self-reported valence/arousal, discrete class labeling, or temporally localized “emotional event” markers for high-fidelity annotation (Yutian et al., 2024, Asif et al., 2022, Freire-Obregón et al., 2024).
• Validation: Protocols include within-subject k-fold (e.g., 10-fold) cross-validation, subject-independent splits (leave-one-subject-out), and stratified splits to ensure robust, generalizable classification statistics (Yutian et al., 2024, Rehman et al., 28 Aug 2025, Kuang et al., 2024).

2. Feature Extraction and Representation Approaches

Accurate EEG emotion classification hinges on the appropriate transformation of non-stationary, spatially distributed EEG signals into feature spaces amenable to statistical or deep learning.

Statistical and Spectral Features

• Band power: Welch’s method or FFT for broadband power (delta, theta, alpha, beta, gamma) is foundational; power integration over canonical bands yields physiologically interpretable features (Rehman et al., 28 Aug 2025, Sateesh et al., 2024).
• Higher-order statistics: Mean, variance, skewness/kurtosis, Hjorth parameters (activity, mobility, complexity), and Shannon entropy further capture amplitude and dynamical characteristics (Rehman et al., 28 Aug 2025, Purwar et al., 31 Jan 2026).
• Differential Entropy (DE): $DE = \frac{1}{2}\ln (2\pi e \sigma^2)$ builds on the variance of band-limited segments, is robust to nonstationarity, and underpins several high-performing pipelines (Sidharth et al., 2023, Ahmed et al., 2022).

Connectivity and Topological Representations

• Phase and amplitude connectivity: Mean Phase Coherence (MPC) and Magnitude Squared Coherence (MSC) are incorporated in upper/lower triangles of synthetic EEG adjacency matrices or “images,” while DE is written to the diagonal, supporting CNN-based classification (Sidharth et al., 2023).
• Asymmetry and AsMap: Pairwise channel differences (asymmetry maps, “AsMap”) in DE capture both lateral and regional emotional markers, suitable as 2D CNN inputs (Ahmed et al., 2022).
• Graph Construction: Explicit learnable neural graph architectures and hierarchical spatial–temporal graphs (HiSTN, MSBAM) encode intra- and inter-regional dependencies, outperforming manually defined connectivity graphs by 10–60 pp in complex tasks (Kuang et al., 2024, Wu et al., 2021). Data-driven edges distinguish functionally relevant neural subnetworks for affect.

Time–Frequency and Deep Learned Features

• Time–frequency tensors: STFT or wavelet decompositions produce 3D arrays for input to 3D CNNs or transformer-based models (Yutian et al., 2024, Karim et al., 6 Feb 2026).
• Automated feature encoding: Deep pipelines supplement hand-crafted features with embeddings obtained from ConvNet, attention, and recurrent modules, which can operate directly on raw, filtered signal windows (Yutian et al., 2024, Karim et al., 6 Feb 2026).

3. Deep Learning Architectures and Model Innovations

Modern EEG emotion classification increasingly exploits deep neural architectures to capture the multiscale, spatiotemporal, and topological complexity of neural signals.

Convolutional and Residual Networks

• 3D CNN and residual blocks: ACPA-ResNet uses stacked 3D convolutions to expand spatial-channel hierarchies, followed by Convolutional Block Attention Modules (CBAM) and fully pre-activated residual blocks. Pre-activation (BN–ReLU–Conv–BN–ReLU–Conv) mitigates gradient vanishing/explosion, while attention modules focus the model on informative electrodes (Yutian et al., 2024).
• Multi-scale temporal CNNs: TSception and MSBAM utilize parallel convolutions at multiple temporal scales, capturing low- and high-frequency emotional EEG components. Asymmetric kernels encode left–right cerebral activity differences crucial for emotion (Ding et al., 2021, Wu et al., 2021).

Recurrent Neural Networks

• LSTM and BiLSTM: Recurrent layers capture long-range EEG sequence context. Stacked LSTM/GRU or hybrid LSTM–GRU models—often deep, with aggressive dropout—are highly effective at leveraging temporal dependencies in multi-class, ordinal, and continuous emotion labeling (Rehman et al., 28 Aug 2025, Sateesh et al., 2024, Asif et al., 2022).
• CNN-LSTM hybrids: 1D/2D CNNs extract spatial/spectral features before passing representations to LSTM blocks, unifying spatial and temporal processing (Asif et al., 2022).

Attention and Transformer Models

• Channel and spatial attention: CBAM (as in ACPA-ResNet) and multi-head attention modules enable the network to dynamically weigh channels and spatial locations, maximizing discriminative power for affective signals (Yutian et al., 2024, Zhu et al., 2023).
• Transformer–CNN fusions: Recent hybrid models embed shallow CNN front-ends for spatial-local feature learning, followed by Transformer encoders (multi-head attention, feedforward) to capture context over windows of time; these yield >91% accuracy in three-class emotion tasks on consumer-grade and reduced-channel EEG (Dolgopolyi et al., 19 Nov 2025, Karim et al., 6 Feb 2026).

Graph- and Hierarchy-Based Models

• HiSTN: Hierarchical Spatial Temporal Networks merge channel-, region-, and global-level graph abstraction, with parameter-efficient (∼1k) models achieving up to 96.8% F1 in 5-way valence tasks (Kuang et al., 2024).
• End-to-end graph learning: Models can simultaneously infer multi-layer dynamic connectivity graphs and perform GNN-based feature extraction, outperforming static-topology baselines by over 60 percentage points in complex video-based emotion tasks (Jang et al., 2019).

4. Labeling Schemes, Validation, and Performance Metrics

EEG-based emotion classifiers employ various labeling and evaluation approaches to accommodate the ambiguities of affect and maximize cross-context and cross-population generalization.

Label Encodings

• Discrete classes: Common taxonomies include binary (low/high valence/arousal), three-way (positive/neutral/negative), four-way (quadrants in valence–arousal space), and fine-grained multi-class (e.g., five classes or more in DREAMER/SEED) (Yutian et al., 2024, Kuang et al., 2024, Dolgopolyi et al., 19 Nov 2025).
• Ordinal and continuum labels: Some frameworks discretize continuous self-report scores into bins or employ soft label smoothing (Gaussian kernel around the actual rating), allowing classifiers to exploit the inherent continuity of emotional experience (Kuang et al., 2024).
• Temporal targeting: Temporally localized event markers (e.g., real-time “emotional event” click labelling) sharply improve accuracy over broad trial-level labeling by focusing learning on emotionally salient epochs (Asif et al., 2022).

Evaluation Protocols

• Cross-validation: Robustness is assessed via k-fold validation (within/provided subject), leave-one-subject-out (LOSO) to test cross-subject generalization, and stratified test splits (Yutian et al., 2024, Rehman et al., 28 Aug 2025).
• Metrics:
  • Classification accuracy is ubiquitous, with top models exceeding 95% for 3–4 classes in subject-dependent settings (Yutian et al., 2024, Sidharth et al., 2023, Vasquez et al., 26 Dec 2025).
  • F1-score (macro or weighted) captures class-balance effects and is essential in imbalanced multi-class regimes (Asif et al., 2022, Rehman et al., 28 Aug 2025).
  • Cohen's κ and confusion matrix analysis support diagnostic interpretation by quantifying agreement and error confusion (e.g., sorrow vs. sadness) (Sidharth et al., 2023, Yutian et al., 2024).
  • Seq2HR: The Sequential Top-2 Hit Rate measures coverage of true scores within top-2 model predictions with small ordinal error, capturing both accuracy and prediction continuity (Kuang et al., 2024).

Comparative Performance

Model/Method	Context	Classes	Accuracy (%)	F1-score (%)	Notes
ACPA-ResNet (Yutian et al., 2024)	8-ch wireless	4	95.1	—	Attention+preact ResNet
LSTM-GRU (Rehman et al., 28 Aug 2025)	GAMEEMO	4	94.5	93.6	Multi-class, hybrid deep
HiSTN (Kuang et al., 2024)	DREAMER	5	—	96.8 (Valence)	78.3 (subj-indep)
ResNet50-Transfer (Sidharth et al., 2023)	SEED	3	93.1–71.6	—	Subject-dep/indep
Dual CNN-Transformer (Dolgopolyi et al., 19 Nov 2025)	SEED+FRA+GER	3	90.8	92.2 (macro)	5 electrodes
BiLSTM (Sateesh et al., 2024)	DEAP	4x9	90.4	—	Sequence-length 256
Random Forest (Vasquez et al., 26 Dec 2025)	Emotiv	3	97.2–76	—	Raw signals, real-time
CNN-LSTM (Asif et al., 2022)	DENS	4	—	96.8	Localized events

A plausible implication is that the best-performing deep learning architectures consistently leverage hybrid spatial-temporal processing, attention mechanisms, and—where possible—graph/region-aware representations, often outperforming classical ML by >5–20 percentage points for moderate to large class cardinality.

5. Challenges, Limitations, and Future Directions

State-of-the-art EEG emotion classification faces multiple challenges intrinsic to affective neuroscience and real-world deployment.

Current Limitations

• Emotion taxonomy depth: Many systems address only 2–4 classes or quadrants; real-world affect is multidimensional and often non-discrete (Yutian et al., 2024).
• Dataset scope: Modest subject numbers (<30 in many studies) and subject-dependent validation protocols limit generality; large-scale, heterogeneous datasets remain rare (Yutian et al., 2024, Kuang et al., 2024).
• Channel and window selection: Trade-offs exist between spatial coverage (more electrodes) and practicality (fewer, for wearable systems), and between temporal localization (short windows for specificity, long for context) (Ahmed et al., 2022, Asif et al., 2022).
• Generalizability: Cross-subject accuracy drops 10–20 pp vs. subject-dependent; transfer and domain-adaptive learning methods remain underexplored (Kuang et al., 2024, Sidharth et al., 2023, Shen et al., 2024).

Research Opportunities

• Richer labels: Future pipelines aim to model continuous valence/arousal, complex or mixed feelings, and even mood change trajectories induced by media and feedback (Yutian et al., 2024, Dolgopolyi et al., 19 Nov 2025).
• Temporal precision: Online/real-time labeling (event-driven annotation) outperforms block-level labels and is advocated for next-generation datasets, improving both ecological and algorithmic validity (Asif et al., 2022).
• Hybrid and lightweight models: Advances in parameter-efficient deep learning (e.g., HiSTN ≈1k params), quantization, and pruning address hardware and latency barriers in edge–BCI applications (Kuang et al., 2024, Dolgopolyi et al., 19 Nov 2025).
• Integrative biosignals: Multimodal fusion with GSR, ECG, and behavioral cues can enhance emotion decoding while also exposing cross-modal redundancy (Lakhan et al., 2018, Rehman et al., 28 Aug 2025).
• Interpretability: Improved graph-based and attention mechanisms, along with feature-importance analysis (e.g., SHAP, saliency maps), provide new means for neuroscientific interpretation and explainability (Kuang et al., 2024, Karim et al., 6 Feb 2026).

6. Neurophysiological Foundations and Interpretability

Modern pipelines are increasingly grounded in neuroscientific principles, aligning algorithmic attention with known affective brain mechanisms.

• Spatial asymmetries: Numerous approaches (e.g., MSBAM, TSception, AsMap) exploit lateralized frontal alpha asymmetry and region-specific patterns (e.g., parietal/frontal shifts for positive/negative affect), with deep models often validating these patterns via saliency analysis (Wu et al., 2021, Ding et al., 2021, Ahmed et al., 2022).
• Spectrotemporal dynamics: Models exploiting dynamic attention (DAEST), temporal multi-scale filtering, or EMD-based spectrograms characterize the rapid state transitions and band-limited phenomena associated with discrete emotions (e.g., high-frequency beta for joy, theta/alpha for anger/sadness) (Shen et al., 2024, Tiwari et al., 2022).
• Connectivity: Data-driven or manually engineered inter-channel connectivity (e.g., phase coherence, graph topologies) is shown to robustly distinguish emotional states, with frequent involvement of frontal-limbic-parietal subnetworks (Jang et al., 2019, Sidharth et al., 2023).

Interpretability remains a challenge for deep and graph-based systems, but advances in layerwise attribution methods, embedding manifold analysis, and explicit physiologically motivated attention modules are enabling finer linkage between model outputs and neural affective substrates.

---

EEG emotion classification has evolved from classical statistical learning on hand-crafted features to sophisticated, attention-guided, graph-based, or transformer-augmented architectures. Present systems routinely surpass 90–95% accuracy in subject-dependent, moderate-class-count settings and are approaching deployable generalization in cross-subject pipelines. Continued progress in labeling granularity, subject diversity, cross-modal integration, and principled model interpretability is anticipated to further entrench EEG emotion decoding as a viable technology for affective computing and neuroadaptive systems.