Adaptive Deep Learning Framework Classifies EEG Emotional States With High Accuracy

Researchers have developed an integrated deep learning architecture that achieves a 98.72 percent classification accuracy in identifying human emotional states from electroencephalogram recordings under a leave-one-subject-out validation protocol. The newly reported framework incorporates Independent Component Analysis for artifact removal, the Wavelet Scattering Transform for multiscale feature extraction, a Graph Attention Network for adaptive channel representation, Adaptive Sparrow Search Optimization for feature selection, and a Graph Transformer Network for final sequence classification. Evaluated on the benchmark DEAP dataset across four discrete emotion quadrants, the end-to-end model demonstrated balanced diagnostic precision of 98.45 percent, recall of 98.31 percent, an F1-score of 98.38 percent, and an area under the receiver operating characteristic curve of 99.12 percent. By testing the framework exclusively on unseen participants through subject-independent cross-validation, the experimental investigation provides a rigorous baseline for objective neural state monitoring in affective computing and intelligent human-computer interaction environments without relying on subject-specific calibration data.

The identification of emotional states through electroencephalography addresses longstanding constraints associated with external affective indicators such as facial expressions, vocal acoustics, and peripheral autonomic measurements. While conventional modalities remain susceptible to conscious suppression, environmental perturbations, and physical movement, electroencephalogram signals directly mirror central nervous system dynamics with millisecond temporal resolution. However, scalp-recorded neuroelectric signals present significant analytical hurdles because they are intrinsically non-stationary, highly nonlinear, and heavily contaminated by physiological and environmental artifacts. Traditional machine learning methodologies frequently depend on handcrafted feature engineering across time, frequency, or statistical domains, which often fails to capture the complex temporal dynamics and spatial interdependencies governing emotional brain states. Furthermore, inter-subject variability and dataset sparsity historically hinder the generalization capacity of computational models when evaluated across different human participants, creating a demand for adaptive deep architectures capable of extracting invariant emotional representations directly from multi-channel neuroelectric measurements.

To overcome signal contamination prior to spectral and spatial modeling, the proposed methodology implements an initial preprocessing stage based on Independent Component Analysis. Because scalp electrodes capture a composite mixture where multiple neural sources and extraneous physiological activities overlap simultaneously, the recorded signal matrix is formulated as the product of an unknown mixing matrix and a set of statistically independent source signals. Through this linear decomposition, the framework isolates underlying neuroelectric generators from pervasive non-neural artifacts, including ocular blinks and muscular contractions. By separating components associated with targeted emotional processing from cognitive control activities and mechanical noise, Independent Component Analysis preserves clean cerebral activity. This noise reduction step ensures that subsequent analytical modules process authentic neural fluctuations rather than high-amplitude physiological distortions, thereby stabilizing downstream representation learning and improving the accuracy of emotional boundary detection across experimental trials.

Following artifact suppression, the framework applies the Wavelet Scattering Transform to convert non-stationary electroencephalogram time series into stable, deformation-resistant mathematical representations. Unlike standard wavelet decompositions that require manual band selection, the scattering transform establishes translation-invariant and multi-scale representations through a cascaded network structure combining wavelet filter convolutions, nonlinear modulus operations, and low-pass averaging functions. Using a Morlet wavelet configuration across two successive layers with a quality factor of eight, the transform processes recordings sampled at 128 Hertz to extract robust first-order and second-order scattering coefficients. The first-order coefficients capture primary envelope modulations through convolution with dilated wavelet filters and averaging filters, while the second-order coefficients characterize higher-order signal interactions and cross-frequency couplings. This signal processing cascade compresses multi-channel temporal traces into comprehensive 256-dimensional feature vectors that retain subtle nonlinear emotional markers while remaining resilient against minor waveform shifts and background acoustic noise.

Spatial relationships and functional interactions among electroencephalogram electrodes are subsequently modeled using a Graph Attention Network. In this topological formulation, individual scalp recording electrodes function as graph nodes whose initial feature embeddings correspond to the 256-dimensional vectors supplied by the Wavelet Scattering Transform, while functional graph edges are defined by Pearson correlation coefficients calculated between channel pairs. The architecture employs two graph attention layers equipped with eight parallel attention heads and a hidden dimension of 128 units to learn expressive spatial embeddings. By dynamically calculating normalized attention coefficients via parameterized weight matrices and leaky rectified linear unit activations, the network selectively emphasizes informative electrode pairs while suppressing irrelevant connections. Incorporating rectified linear unit activation functions and a dropout probability of 0.5 prevents structural overfitting, enabling the system to map functional connectivity patterns across emotionally relevant cerebral regions during subject-independent classification tasks.

To refine the dimensional complexity of the learned graph embeddings, the architecture incorporates Adaptive Sparrow Search Optimization to execute automated feature selection and hyperparameter tuning. Inspired by the foraging dynamics and antipredation responses observed in sparrow flocks, the optimization routine divides candidate solutions into producers that execute wide-area global exploration, scroungers that follow successful producers to perform localized exploitation, and vigilantes that detect suboptimal clustering to avoid entrapment in local optima. The algorithm integrates an adaptive weight adjustment mechanism that systematically reduces search weights across consecutive iterations, balancing exploratory capability with rapid convergence speed. Guided by a specialized fitness function designed to maximize classification accuracy while penalizing redundant feature counts according to explicit weighting coefficients, the optimizer systematically ranks feature subsets evaluated by the classifier, discarding uninformative variables and boosting overall computational efficiency without sacrificing discriminative precision.

The sequence of optimized spatial embeddings is subsequently classified using a two-layer Graph Transformer Network designed to capture both localized transitions and long-range temporal dependencies. The electroencephalogram trials are segmented into uniform time windows from which sequential graph representations are derived and indexed using positional encoding to preserve temporal order. Within the transformer blocks, multi-head self-attention mechanisms operate across a hidden dimensionality of 128 to evaluate inter-channel relationships and structural shifts across consecutive time windows. The resulting node representations across all recording channels undergo global mean pooling to construct a unified emotional graph representation vector. This pooled vector is projected through a fully connected classification layer to produce raw output logits, which are converted into normalized class probabilities using the Softmax activation function. Model optimization is conducted through backpropagation using the categorical cross-entropy loss function paired with the Adam optimizer operating at an initial learning rate of 0.001.

The computational framework was experimentally implemented and evaluated using the publicly accessible DEAP dataset, a benchmark database comprising multi-channel recordings from 32 healthy participants divided equally between males and females. During data acquisition, subjects viewed 40 one-minute music video presentations structured to elicit diverse affective states, after which they recorded self-assessment ratings for valence, arousal, dominance, and liking on a nine-point Likert scale. For computational modeling, 32 electroencephalogram channels downsampled to 128 Hertz were classified into four discrete emotional quadrants: High Valence-High Arousal, High Valence-Low Arousal, Low Valence-High Arousal, and Low Valence-Low Arousal, utilizing a threshold split positioned at the midpoint score of 5.0. Experimental execution was conducted in a Python 3.11 environment utilizing the PyTorch library on a dedicated computing workstation configured with an Intel Core i7 central processing unit, 32 gigabytes of random-access memory, and an NVIDIA RTX 3060 graphics processing unit with 12 gigabytes of onboard memory, applying a uniform batch size of 32 across 20 training epochs.

Performance trajectories across the 20 training epochs revealed steady convergence behavior, starting from initial training and validation accuracies of approximately 82 percent and 80 percent, and initial loss values of 0.52 and 0.56, respectively. Between epochs 1 and 10, rapid parametric optimization drove sharp accuracy improvements, after which metric curves leveled into stable asymptotes where training accuracy reached nearly 99.9 percent and validation accuracy attained 99.4 percent, with final training and validation losses settling at 0.01 and 0.03. Comparative baseline evaluations conducted under identical leave-one-subject-out protocols demonstrated that the proposed model substantially outperformed standard convolutional networks, recurrent long short-term memory networks, hybrid convolutional-recurrent architectures, and conventional graph convolutional networks. In parallel, systematic ablation experiments confirmed the discrete utility of each architectural module: combining the scattering transform with the graph transformer yielded 95.23 percent accuracy, integrating the graph attention network elevated accuracy to 96.25 percent, incorporating sparrow search feature optimization achieved 97.06 percent, and the complete unified framework reached the maximum observed accuracy of 98.72 percent.

The observed experimental results demonstrate that combining multi-scale wavelet representations, spatial graph attention, swarm-based feature selection, and temporal graph transformers provides structural stability against signal noise and missing electrode channels. The confusion matrix verified that most sample trials were correctly categorized along the principal diagonal, with negligible misclassifications restricted to adjacent emotional quadrants. However, the authors emphasize that their current validation remains constrained to the 32-subject cohort of the DEAP database under controlled experimental parameters. Future research initiatives must focus on validating the architecture across alternative public databases, such as SEED, DREAMER, and AMIGOS, while systematically evaluating cross-dataset generalization, cross-device hardware adaptations, and performance degradation under simulated electrode loss. Additionally, future experimental extensions will include the formal reporting of subject-wise mean performance metrics, standard deviations, and statistical confidence intervals to establish clinical and operational boundaries for wearable neurotechnology applications.

Subject of Research: Electroencephalogram-based emotion recognition using an adaptive deep learning framework

Article Title: Recognition of emotional state using EEG signals using an adaptive deep learning model

Article References: Alagirisamy, M., Nataraj, C., Selvaperumal, S. K., Sivanesan, S., Sonavale, A., & Jadhav, J. (2026). Recognition of emotional state using EEG signals using an adaptive deep learning model. Discover Artificial Intelligence, 6(1), Article 1792. https://doi.org/10.1007/s44163-026-02327-4

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02327-4

Keywords: affective computing, electroencephalography, deep learning, graph attention network, graph transformer network, wavelet scattering transform, adaptive sparrow search optimization, DEAP dataset, Recognition, emotional, state, signals