@Article{jqc.2026.086882,
AUTHOR = {Reon Yoshida, Keiko Ono, Kentaro Ohki, Takuya Futagami},
TITLE = {EEG-Based Emotion Recognition Using Deep Quantum Features},
JOURNAL = {Journal of Quantum Computing},
VOLUME = {8},
YEAR = {2026},
NUMBER = {1},
PAGES = {123--144},
URL = {http://www.techscience.com/jqc/v8n1/68716},
ISSN = {2579-0145},
ABSTRACT = {Quantum machine learning (QML) has attracted significant attention for its potential to accelerate computation and improve efficiency, particularly through quantum feature maps that may enable the separation of data not linearly separable in classical spaces. Although this capability remains largely theoretical, it represents a promising direction for addressing complex learning tasks. However, current quantum devices suffer from low error tolerance and a limited number of qubits, which has spurred interest in hybrid quantum–classical approaches. One such application is EEG-based emotion recognition, which involves complex, nonlinear signals and substantial inter-subject variability. While prior studies have applied QML to EEG analysis—using QSVMs on handcrafted features, sequential quantum-then-classical pipelines on band-power attributes, or fully quantum hybrid models for brain–computer interfacing—none of these approaches integrate high-level convolutional representations with entanglement-based quantum feature maps in a parallel fusion architecture, nor do they investigate subject-wise adaptive contribution of quantum features. To address these gaps, this study proposes a hybrid architecture that integrates classical and quantum computing by leveraging <i>deep quantum features</i>. These features are synthesized by combining deep representations extracted via a Convolutional Neural Network (CNN) with quantum features generated through a ZZFeatureMap-based quantum circuit using angle encoding, in which each of four qubits is semantically aligned with a canonical EEG frequency band (<math id="mml-ieqn-1"><mi>θ</mi></math>, <math id="mml-ieqn-2"><mi>α</mi></math>, <math id="mml-ieqn-3"><mi>β</mi></math>, <math id="mml-ieqn-4"><mi>γ</mi></math>). The two branches are concatenated and processed by a two-layer multilayer perceptron for binary classification of Valence and Arousal on the DEAP dataset. Across ten independent training runs, the proposed model achieved average classification accuracies of 79.6% for Valence and 81.5% for Arousal. Based on the normality of the paired subject-level differences, two-sided paired <i>t</i>-tests or Wilcoxon signed-rank tests showed significant improvements over the standalone classical and quantum models (<math id="mml-ieqn-5"><mi>p</mi><mo>&lt;</mo><mn>0.05</mn></math>) under the same segment-level subject-dependent protocol. Additional robustness and ablation analyses further showed that the same ordering, Proposed <math id="mml-ieqn-6"><mo>&gt;</mo></math> Classical <math id="mml-ieqn-7"><mo>&gt;</mo></math> Quantum, was preserved under repeated-seed evaluation, and that replacing the quantum branch with shape-matched random noise substantially degraded performance. Furthermore, interpretability analysis using LIME showed subject-dependent use of both feature branches, with the quantum-derived branch receiving greater importance in many cases. This result supports the usefulness of the quantum-derived representation as complementary predictive information, although it should not be interpreted as causal evidence of a uniquely quantum advantage. Together, the parallel CNN–quantum fusion design, the band-aligned qubit encoding, and the subject-wise interpretability analysis distinguish this work from existing QML-EEG studies, and highlight the efficacy of classical–quantum hybrid approaches for complex biomedical signal analysis.},
DOI = {10.32604/jqc.2026.086882}
}