Use of cognitive load measurements to design a new architecture of intelligent learning systems

Amin Zammouri, Abdelaziz Aït Moussa, Sylvain ChevallierExpert Systems with Applications 2024 · DOI

Published in Expert Systems with Applications, this research introduces a novel approach to Intelligent Learning Systems (ILS) by integrating real-time brain activity monitoring. By measuring a learner's mental effort alongside their task performance, the system can more accurately assess true mastery and adapt pedagogical content to prevent cognitive overload or disengagement.

The Problem: The "Blindness" of Traditional Tutoring

Traditional Intelligent Tutoring Systems (ITS) typically adapt content based solely on performance metrics—if a student provides a correct answer, the difficulty increases. However, this approach ignores the internal cost of success. A student might achieve a correct result while operating at their absolute cognitive limit (high load), or fail a simple task due to boredom (low load).

While physiological markers like heart rate variability (HRV) or skin conductance (GSR) can track general stress, they often lack the temporal resolution required for complex learning tasks. This paper bridges that gap by utilizing Brain-Computer Interfaces (BCI) to provide a nuanced, real-time metric of "mental effort."

Proposed Multi-Agent Architecture

The researchers developed a multi-agent architecture designed to process neurophysiological data in tandem with educational progress. The system is divided into four primary modules: Interface, Pedagogical, Learner, and Domain.

The multi-agent architecture of the proposed Intelligent Learning System, featuring specialized agents for brain load estimation and pedagogical planning.

Key technical components include:

  • Brain Load Estimation Agent: Instead of complex, pre-trained classifiers, the system uses Standardized Euclidean Distance (SED). It calculates the distance between the learner's current Power Spectral Density (PSD) in the Theta (4-7 Hz) and Alpha (8-11 Hz) bands against a "quasi-rest" baseline.
  • Fuzzy Logic Evaluation: The system employs a fuzzy logic controller that synthesizes two inputs: response accuracy ($Corr_{out}$) and estimated cognitive load ($ASED^i$). This produces a "Knowledge Level" that reflects both outcome and effort.

Methods and Neuro-Indices

The study involved 30 male university students across three protocols: Raven’s Progressive Matrices (to validate rest vs. task states), an auditory N-Back test (to identify brain regions modulating during memory tasks), and a custom mathematical logic test.

EEG electrode placement map (10-20 system) highlighting the occipital regions (green) found to be highly sensitive to cognitive load.

One of the study's most significant findings was the Occipital Surprise. While most cognitive load research focuses on the frontal lobe for working memory, this study found that the occipital lobe (the back of the brain) showed significant Alpha and Theta modulations even during non-visual tasks. This suggests that posterior brain regions are more involved in general cognitive load than previously emphasized in ILS literature.

Key Results and Insights

  • Sensitivity to Difficulty: The $ASED^i$ (Average SED) metric showed a statistically significant increase as task difficulty moved across levels, confirming its utility as a load indicator.
  • Validation: The proposed method achieved a Cohen’s Kappa coefficient of 0.63 (on electrode O2), indicating "strong agreement" with established neuro-indices like Event-Related Desynchronization/Synchronization (ERD/ERS).
  • Nuanced Assessment: The fuzzy logic system matched traditional scoring 90% of the time. Crucially, in the remaining 10%, it provided "course corrections" where high mental effort suggested the learner had not yet reached true fluency, despite getting the answers right.

Limitations and Future Directions

The study was limited to a demographic of 30 male university students; brain rhythms can vary significantly across ages and genders. Furthermore, the experiments were conducted in controlled settings to minimize EEG noise.

Future work aims to apply this architecture to clinical populations, such as children with dyslexia or patients with cerebral palsy, to create more sensitive rehabilitation tools. The researchers also plan to incorporate emotional stress factors into the model to create a more holistic view of the learner's state.

Summary updated

BibTeX
@article{zammouri2024use,
  author = {Amin Zammouri and Abdelaziz Aït Moussa and Sylvain Chevallier},
  title = {{Use of cognitive load measurements to design a new architecture of intelligent learning systems}},
  journal = {Expert Systems with Applications},
  year = {2024},
  doi = {10.1016/j.eswa.2023.121253},
}
Figure