Subject-Wise Cognitive Load Detection Using Time–Frequency EEG and Bi-LSTM

• Published in: Arabian journal for science and engineering (2011) Vol. 49; no. 3; pp. 4445 - 4457
• Main Authors: Yedukondalu, Jammisetty, Sharma, Diksha, Sharma, Lakhan Dev
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.03.2024; Springer Nature B.V
• Subjects: Cognitive load; Electroencephalography; Engineering; Humanities and Social Sciences; multidisciplinary; Research Article-Computer Engineering and Computer Science; Science; Time-frequency analysis; Bi-LSTM; Subject-wise; Cognitive load; T–F methods; EEG
• ISSN: 2193-567X; 1319-8025; 2191-4281
• DOI: 10.1007/s13369-023-08494-1

Cognitive load detection using electroencephalogram (EEG) signals is a technique employed to understand and measure the mental workload or cognitive demands placed on an individual while performing a task. EEG is a noninvasive method that records fluctuations in brain activity at different cognitive load levels. The publicly available multi-arithmetic task EEG dataset was used. This study introduces a novel approach to detecting cognitive load by utilizing both the 1D-EEG signal and its various time–frequency (T–F) representations as 2D images. The signal underwent preprocessing, including artifact-free segmentation using filters and subsequent normalization, before being fed into a bidirectional long short-term memory (Bi-LSTM) model with different optimizers for classification. It was trained and fine-tuned to achieve high accuracy. Remarkably, our proposed method demonstrates promising performance even with short EEG segments as 4 s. Through 10-fold cross-validation, we achieved an accuracy (Ac%) of 99.55 and 99.88 using 5:5 and 8:2 data splits, respectively. Furthermore, this manuscript includes subject-wise cognitive load detection, providing valuable insights into individual cognitive processes. This approach enables targeted interventions, performance optimization, and mental health monitoring across various domains. For 36 subjects, an average Ac% of 85.22 was attained. Notably, the spectrogram T–F conversion-based 2D image, coupled with a Bi-LSTM classifier and Adam optimizer, outperformed previous state-of-the-art techniques in terms of evaluation metrics.