Mind Wandering is Not Monolithic

Code repository for "Mind Wandering is Not Monolithic: Latent Profiles of Mind Wandering Dimensions Reveal Distinct Gaze Signatures Across Environmental Conditions" — for the EYEMW Dataset Competition (deadline: April 15, 2026).

Research Question

Do different dimensions of mind wandering form distinct, empirically separable profiles with unique gaze signatures, and do these profiles remain stable across environmental conditions?

Key Findings

• MW is not monolithic: Low-to-moderate correlations between MW dimensions (max |r| = 0.51)
• 5 MW profiles identified via Latent Profile Analysis in video tasks (n=8,167) and listening/reading tasks (n=2,674)
• Gaze discrimination: Webcam-based gaze features achieve AUROC = 0.59 for multi-profile classification
• Environmental robustness: Feature importance rankings are stable across lighting (rho=0.71-0.86) and device types (rho=0.88)

Repository Structure

EYEMW/
├── database/                    # EYEMW dataset (not included; obtain from eyemindwander.com)
├── scripts/
│   ├── phase1_data_preparation.py   # Data loading, harmonization, eligibility
│   ├── phase2_lpa.py                # Latent Profile Analysis + bootstrap validation
│   ├── phase3_gaze_discrimination.py # RF classifiers, feature importance, cross-task
│   ├── phase4_slicing_analysis.py   # Environmental robustness slicing
│   ├── phase5_figures.py            # Publication figures and summary tables
│   └── run_all.py                   # Execute full pipeline
├── results/                     # CSV/JSON/parquet outputs
├── figures/                     # PNG figures
├── results_explorer.ipynb       # Interactive walkthrough of all results
├── requirements.txt
└── README.md

Setup

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Running

# Full pipeline
source .venv/bin/activate
python scripts/run_all.py

# Or individual phases
python scripts/phase1_data_preparation.py
python scripts/phase2_lpa.py
python scripts/phase3_gaze_discrimination.py
python scripts/phase4_slicing_analysis.py
python scripts/phase5_figures.py

Methods Summary

1. Data Harmonization: Binary probes kept as 0/1; Likert scales min-max normalized to 0-1; gaze features z-scored within study
2. Latent Profile Analysis: Gaussian Mixture Models (k=2-6), BIC-selected, bootstrap-validated (1000 resamples)
3. Gaze Discrimination: Random Forest (500 trees, balanced weights, 5-fold CV) with multiclass AUROC
4. Environmental Slicing: Within-slice AUROC and feature importance rank correlations across lighting, device, and setting conditions