EEG-based Prediction of Cognitive Load in Intelligence Tests

Friedman, Nir; Fekete, Tomer; Gal, Ya’akov (Kobi); Shriki, Oren · 2019 · Crossref

DOI: 10.1101/539486

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This study addresses the challenge of quantifying cognitive load during complex problem-solving tasks using electroencephalography (EEG). While previous research often relied on discrete low/high load classifications or simple spectral features, this work aims to predict cognitive workload as a continuous variable. The authors utilized the Advanced Progressive Matrices (Raven’s Matrices), a validated intelligence test with rigorously defined difficulty levels, as a proxy for increasing cognitive load. The motivation stems from the need for accurate, real-time monitoring of mental workload in applications such as e-learning, military training, and pilot monitoring. The experimental design involved 47 participants who completed 36 Raven’s Matrices problems of increasing difficulty while their EEG was recorded using a 64-channel system. Data were preprocessed to remove artifacts, and three categories of features were extracted: power spectrum metrics (delta to gamma bands), neural complexity measures (Lempel-Ziv complexity, Multi-Scale Entropy, and Detrended Fluctuation Analysis), and network connectivity metrics derived from complex network analysis. These features, along with basic demographic and response-time data, served as inputs for four machine learning algorithms: Linear Regression, Random Forest, XGBoost, and Artificial Neural Networks (ANN). The models were trained to predict the specific difficulty level of each problem. The results demonstrated that cognitive load could be predicted with significant accuracy. XGBoost outperformed all other algorithms, achieving a Pearson correlation coefficient ($r^2$) of 0.67 and a Spearman correlation of 0.81 between predicted and actual difficulty levels. Feature analysis revealed that combining power spectrum and connectivity metrics yielded the best performance, while adding complexity features did not significantly improve predictions. Notably, the study found that prediction quality remained robust even when reducing the number of EEG channels; using only the 12 most informative electrodes achieved an $r^2$ of 0.70. Furthermore, individualizing the ANN model by tuning the output layer for each subject significantly improved prediction accuracy compared to a general model ($p = 0.001$). The significance of these findings lies in the demonstration that continuous cognitive load can be reliably inferred from EEG data using machine learning, even with a reduced number of electrodes. This supports the feasibility of deploying portable, low-channel EEG systems for real-time cognitive monitoring in practical settings. The study highlights that connectivity features add valuable information beyond spectral power, and that personalized models offer superior performance over general ones, paving the way for adaptive learning systems and optimized human-machine interfaces.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-nvidia on 2026-08-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• neuro workload indices
• workload measurement