The determination of human creative thinking by employing machine learning classification on EEG signals

Traditional creativity assessments are limited by subjectivity and high labor costs. Although machine learning (ML) offers objective alternatives, its application to EEG-based creativity evaluation remains scarce. This study aimed to classify high and low creative thinking from EEG signals using ML.

One hundred forty participants completed the Alternative Uses Task during EEG recording. Three independent raters (none were authors) evaluated response originality using the Consensus Assessment Technique on a 1-to-5 scale; mean scores were dichotomized at the median into high- and low-creativity labels (996 and 1,096 trials, respectively, from 2,092 valid trials). Classification features included alpha-band Power Spectral Density (PSD), Approximate Entropy, Sample Entropy, and a combined feature set. Six classifiers-Support Vector Machine (SVM), Quadratic Discriminant Analysis (QDA), Logistic Regression (LogR), Decision Tree (DT), XGBoost, and LightGBM-were trained and evaluated using a 10-fold cross-validation strategy. To prevent subject-level information leakage, a Leave-One-Subject-Out (LOSO) validation was additionally conducted.

All six classifiers effectively distinguished creativity levels. Under 10-fold cross-validation, SVM achieved optimal performance using Approximate Entropy or Sample Entropy (F1-score = 90.5%; accuracy = 89.8%). The combined feature set yielded comparable results. LOSO validation confirmed generalizability to unseen individuals, with SVM attaining F1-scores of 82.4% (Approximate Entropy) and 82.1% (Sample Entropy). Entropy-based features consistently outperformed alpha PSD.

ML effectively classifies creativity from EEG signals. The superior performance of entropy features, supported by both trial-level and subject-independent validation, highlights the robustness of the proposed approach and its potential for developing objective, scalable creativity assessment tools.