Machine-learning applications for differentiation across states and stages of creative thinking based on time-series and time–frequency features of electroencephalography and event-related potential signals

Abstract

BACKGROUND: The study presents ML approaches for creative tasks states/stages differentiation using "test-control" approach. Control tasks were considered as initial stages of creative task fulfillment. The comparative study using both: time-series, time-frequency maps (wavelet transformation) classifications at different stages of creating - endings to well-known proverbs (PROVERBS), plot stories (STORIES) or visual images/painting (viART) was carried on. AIM: Compare and choose machine learning classification approaches for EEG signal separation at different stages/states of creative tasks fulfillment. METHODS: 22 persons participated in PROVERBS (ERP paradigm), 15 in the STORIES creation tasks. The case study of artistic creativity was carried out with professional artist (divergent thinking paradigm). We used both linear and CNN classifiers. EEG data was previously ICA corrected conversed to CSD. Continuous EEGs were divided into 4s intervals; and 1500 ms after stimulus presentation were used in ERPs. The EEG/ERP time-frequency maps (Morlet) for 3-30 Hz were generated for 4s intervals with 100 ms increment (continuous EEGs STORIES, viART) or for 1500 ms after stimulus presentation (ERPs PROVERBS) and consisted of combined images (224x224 pix) for frontal (Fz) and parietal (Pz) zones. Images classification was carried out with modified CNN (ResNet50 architecture). RESULTS: Four classes offline accuracy in STORY creation task (inventing a plot, continuation of plot, picture description,background - open eyes) was up to 96.4% [8.3 SD] with ResNet architectures (ResNet50, ResNet18). Three states of artists creative painting (viART) - (background with open eyes, painting on canvas, viewing the painting) gave - 86.94% (Kernel Naive Bayes) and 98.2% for CNN. While trained and tested samples were given for the CNN in consequence order (neuro interface mode) the accuracy in average diminished to 70% [11%SD]. In ERP paradigm (PROVERBS)- 3 class accuracy (creation of new" ending, synonim naming, remembering the known ending) was 80.5% [8.7SD] for CSP with rSVM, while CNN gave - 43.2% [8.8 SD]. CONCLUSION: At the moment, the use of convolutional neural networks has shown a relatively better result for the classification of "continuous", long-term states of creative activity. While estimating fast "transients" were more effective at classifying "time-series with spatial filtering.