Analysis of single channel electroencephalographic signals for visual creativity: A pilot study. (May 2022)

• Record Type: Journal Article
• Authors: Gopan K, Gopika; Reddy, S.V.R. Aditya; Rao, Madhav; Sinha, Neelam
• Is Part Of: Biomedical signal processing and control. Volume 75(2022)
• Publication Date: 2022-05
• Subjects: EEG -- Visual creativity -- Creativity assessment -- Classification -- Neurosky; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.103542
• Languages: English
• ISSNs: 1746-8094

Abstract

Graphical abstract: Highlights: The work attempts to utilize characteristic differences in EEG to distinguish varying levels of visual creativity. Four different frameworks are analyzed namely. (1) Chaos, (2) Distribution, (3) Statistical and (4) LSTM. In 4-class classification, the chaos framework results in the peak performance. This study sheds light on the potential of using commercially available EEG headset to assess visual creativity. Abstract: Analysis of creativity has important clinical applications in studying neurological disorders. Among the available tools, the simplest and the most portable device is the EEG headset. In this study we explore the potential of a single channel EEG to assess varying levels of Visual Creativity via sketching. Three tasks requiring different levels of creativity are designed: Max Creative(TMC)-sketching, Less Creative(TLC)-repetitive geometric patterns, and Nil Creative(TNC)-tally marks. Normalized and artifact-filtered EEG signals are used for analyses. Three different types of analysis are carried out using the features extracted from three different paradigms: (1) Chaos Analysis, (2) Distribution Analysis, and (3) Statistical Analysis. The 4-class classification scenario (Rest vs TMC vs TLC vs TNC) is compared with LSTM which is the popular technique for 1D signal. Mean accuracy is reported over 5-fold cross-validation over 10 runs. Among various feature and classifier combinations, peak performances are seen in the following scenarios. In 4-class classification, chaos features result in a peak mean accuracy of 52% while other features show significantly less accuracy. LSTM for 4-class classification results in only 38% accuracy in the current setting. When levels of creativity(TMC vs TLC vs TNC) are analyzed, Hjorth mobility and complexity, LLE and distribution features result in 45% accuracy. As expected, it is observed that Rest(R) is distinguished from Non-Rest with a mean accuracy > 90% using Standard deviation, MAD and LLE, individually. However, distribution-based features result in 69-74% accuracy. It is observed that chaos analysis results in a higher accuracy for the classes considered.