Decoding Study-Independent Mind-Wandering from EEG using Convolutional Neural Networks

Jin, Christina Yi; Borst, Jelmer P.; van Vugt, Marieke K. · 2020 · Crossref

DOI: 10.1101/2020.12.08.416040

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 decoding mind-wandering from electroencephalography (EEG) data with high generalizability across different experimental contexts. Previous research using support vector machines (SVMs) and hand-crafted features often achieved a performance ceiling of approximately 60% accuracy when generalizing across participants or tasks. The authors hypothesized that convolutional neural networks (CNNs), which can learn directly from raw EEG data with minimal preprocessing, might improve decoding accuracy and robustness. The primary goal was to determine if CNNs could successfully classify mind-wandering states in an inter-subject, across-study setting, thereby moving toward a study-independent biomarker for this mental state. The researchers utilized two independent EEG datasets: a training dataset (Dataset A) from Jin et al. (2019) and a testing dataset (Dataset B) from Jin et al. (2020). Both datasets involved participants performing sustained attention to response and visual search tasks, with mind-wandering states identified via probe questions. EEG data were preprocessed to include 32 overlapping channels, with epochs extracted from -400 ms to 1000 ms relative to stimulus onset. The study compared four input types: raw EEG, band-frequency power, single-trial event-related potentials (stERP), and inter-site phase clustering (ISPC) connectivity matrices. Experiment 1 evaluated three CNN architectures on intra-subject models to identify the optimal structure and input type. Experiment 2 tested the best-performing model (M2) on inter-subject predictions, training on Dataset A and testing on Dataset B. A stacking model combined outputs from the top-performing channels to enhance generalization. In Experiment 1, the simple CNN architecture (M2) with two convolutional layers outperformed deeper or narrower variants. Raw EEG input yielded the highest validation accuracy (71%), followed by stERP (69%), ISPC (62%), and power (57%). Clustering analysis revealed that high validation accuracies in some cases resulted from classification bias toward the majority class rather than genuine detection of minority mind-wandering instances. In Experiment 2, the inter-subject model achieved an accuracy of 68% when predicting mind-wandering in the independent testing dataset. This result verified the generalizability of the CNN approach across different studies, participants, and task designs. The study also found that training on subsets of more balanced datasets improved performance on minority cases, suggesting that class imbalance significantly impacts model reliability. The findings demonstrate that CNNs can effectively decode mind-wandering from raw EEG data with better generalizability than previous methods. The ability to achieve 68% accuracy across independent studies indicates that CNNs can capture robust neural signatures of mind-wandering that transcend specific experimental contexts. The authors conclude that raw EEG is the superior input for these models and recommend that future EEG-machine learning studies explicitly report dataset balancing status and class-specific accuracies to ensure valid performance evaluation. This work supports the potential for developing real-time, study-independent tools for monitoring attentional states.

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