On EEG Preprocessing Role in Deep Learning Effectiveness for Mental Workload Classification

Kingphai, Kunjira; Moshfeghi, Yashar · 2021 · Crossref

DOI: 10.1007/978-3-030-91408-0_6

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 lack of standardized preprocessing protocols in deep learning applications for mental workload (MWL) classification using Electroencephalogram (EEG) signals. While deep learning models show promise in detecting MWL, inconsistent preprocessing techniques across studies hinder the comparability of results and obscure the true effectiveness of these models. The authors investigate how specific, automated preprocessing techniques—high-pass filtering, the ADJUST algorithm for artifact removal, and re-referencing—affect the performance of state-of-the-art deep learning models. By focusing on automated methods, the study aims to eliminate human bias and establish a uniform methodological framework for EEG-based MWL analysis. The researchers utilized the publicly available STEW dataset, comprising EEG signals from 48 male participants recorded during resting states and multitasking activities. They designed four experimental scenarios: raw data (no preprocessing), high-pass filtering only, high-pass filtering combined with the ADJUST algorithm, and the full pipeline including re-referencing. Features such as power spectral density (alpha and theta bands), skewness, kurtosis, approximate entropy, and Hurst exponent were extracted from 14 channels. These features were input into three deep learning architectures: Stacked LSTM, Bidirectional LSTM (BLSTM), and BLSTM-LSTM. The models were evaluated on two tasks: binary classification (resting vs. working) and multiclass classification (low, moderate, and high MWL). Performance was measured using sensitivity, specificity, precision, accuracy, false acceptance rate, and false rejection rate, with statistical significance assessed via p-values. The results demonstrate that preprocessing significantly enhances model performance across all architectures. The ADJUST algorithm had the most substantial impact on improving classification metrics compared to other steps. The full preprocessing pipeline (Scenario 4) yielded the highest accuracy and lowest error rates for all models. In Task 1, the BLSTM-LSTM model achieved an accuracy of 89.44% with the full pipeline, compared to 81.70% with raw data. In Task 2, the BLSTM-LSTM model reached 91.15% accuracy with full preprocessing, a notable improvement from 79.90% with raw data. Statistical analysis confirmed that these improvements were significant (p < 0.05). The study also found that while more complex models generally performed better, the benefits of preprocessing were consistent across all model types. The findings underscore the critical role of standardized, automated preprocessing in maximizing the effectiveness of deep learning models for MWL classification. The study concludes that applying a high-pass filter, the ADJUST algorithm, and re-referencing creates a robust pipeline that significantly reduces noise and improves classification accuracy. This work provides a foundational step toward establishing uniform methodological standards in the field, enabling more reliable comparisons between different studies and facilitating the development of automated, real-time MWL monitoring systems.

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