Deep Convolutional Neural Network for Automated Detection of Mind Wandering using EEG Signals

Hosseini, Seyedroohollah; Guo, Xuan · 2019 · Crossref

DOI: 10.1145/3307339.3342176

archive: archived pipeline: cataloged verified

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

Summary

This paper addresses the challenge of automatically detecting mind wandering (MW), a ubiquitous phenomenon where attention shifts from task-related activities to task-unrelated thoughts. MW is associated with decreased performance, errors, and emotional disorders, creating a need for intelligent interfaces that can detect and reorient attention. While previous research has relied on oculometric measures or fMRI, and some studies have explored EEG signals using traditional feature extraction, this work proposes the first deep learning model to detect MW using only raw EEG data. The authors aim to develop a data-driven technique that accurately and efficiently identifies MW episodes without manual feature engineering. The study utilizes EEG data collected from two participants (one female, age 25; one male, age 31) performing a breath-counting task. Participants pressed a button when they realized they had lost track of their count, marking the onset of MW. The authors employed a cropped training strategy, extracting 8-second time windows from the continuous EEG recordings. Specifically, they selected 475 samples of focusing state (FS) and 475 samples of MW, resulting in 950 total samples. The data underwent bandwidth filtering (0.5–50 Hz) and Z-score normalization. The proposed model is a channel-wise deep convolutional neural network (CNN) with 12 layers, including four convolution-max-pooling blocks for feature extraction and three fully connected layers for classification. The architecture processes 64-channel EEG inputs represented as 2-D arrays, using temporal and spatial convolutions to capture signal correlations. The model was trained using the Adam optimizer with cross-entropy loss and evaluated using ten-fold cross-validation. Experimental results demonstrate that the 8-second time window yielded the best performance, achieving a classification accuracy of 91.78%, sensitivity of 92.84%, and specificity of 90.73%. This outperformed models using 5-second (86.63% accuracy) and 2-second (78.52% accuracy) windows. However, the study revealed significant individual variation in model performance. When the model was trained on one participant’s data and tested on the other, accuracy dropped substantially to 67.63% and 65.26%, respectively. In contrast, training and testing on mixed data from both participants resulted in 81.84% accuracy. These findings indicate that MW and focusing states possess unique, subject-specific EEG features that hinder generalization across individuals. The significance of this work lies in demonstrating the viability of deep CNNs for automated MW detection using raw EEG signals, eliminating the need for explicit feature extraction steps. The high accuracy within the combined dataset suggests the model’s potential for real-time monitoring applications. However, the poor cross-subject generalization highlights a critical limitation: current models are sensitive to individual physiological differences. The authors conclude that future research must focus on acquiring diverse datasets to improve generalization and potentially incorporate subject-specific adaptations to make automated MW detection robust for broader populations.

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.

• drowsiness detection algorithms