Machine-learning applications for differentiation across states/stages of creative thinking based on time-series and time-frequency features of EEG/ERP signals

Abstract

BACKGROUND: The study presents machine-learning (ML) classification approaches for the state/stage differentiation of creative tasks using the “test-control” approach. The control tasks were considered as the initial stages of the creative activity. Time-series and time-frequency electroencephalography (EEG) data analyses were employed in three divergent thinking tasks: 1) creating endings to well-known proverbs (“PROVERBS”, event-related potential [ERP] paradigm); 2) creating stories (“STORIES”, continuous EEG); 3) free creative painting (“viART”, continuous EEG). AIM: To compare and select effective ML classification approaches for EEG signal separation at different stages or states of creative task performance. METHODS: In this study, 22 individuals participated in the “PROVERBS” (ERP paradigm), 15 in the “STORIES”, and 1 (a longitudinal case study) in the “viART” tasks. Linear and convolutional neural network (CNN) classifiers were used. EEG data were previous artifacts corrected and converted to current source density (CSD). Continuous EEGs were divided into 4-s intervals and 1500 ms after stimulus presentation, were used in ERPs. The EEG/ERP time-frequency maps (Morlet wavelet transformation) for 3–30 Hz were generated for 4-s intervals with 100 ms shift (continuous EEGs in “STORIES” and “viART”) or for 1500 ms after stimulus presentation (ERPs in “PROVERBS”) and consisted of combined images (224×224 px) for frontal (Fz) and parietal (Pz) brain zones. Image classification was carried out using the modified CNN (ResNet50, ResNet18 architectures). RESULTS: The offline classification accuracy of the four-class system (description of a picture, inventing a story plot, continuation of story’s plot, and background with open eyes) in the “STORY” creation task was up to 96.4% [±8.3 SD] with ResNet architectures (ResNet50 and ResNet18). The accuracy of the three states discrimination of the artists’ creative painting (resting state with open eyes, painting on canvas, and viewing the painting) was 86.94% for kernel naive bayes and 98.2% for CNN. For the trained and tested samples given for the CNN in consecutive order (neurointerface mode), the accuracy diminished to 70.0% [11% SD] on average. In the ERP paradigm “PROVERBS”, the classification accuracy of the three-class system (creation of “new” ending, naming of semantic synonym, and remembering of the known ending) was 80.5% [±8.7 SD] for the common spatial pattern, followed by rSVM (radial kernel basis support vector machine), compared with 43.2% [±8.8 SD] for CNN. CONCLUSION: The use of CNNs allowed better classifying of “continuous” long-term states of creative activity. In fast “transient processes” such as ERP, time-series classifiers with spatial filtering proved to be more efficient.