tdnathmlenthusiast/cognitive-fatigue-detector

Cognitive Fatigue Detector

Two-stage transfer learning pipeline for real-time fatigue detection from eye-tracking signals.

Architecture

SharedEncoder: Linear(in→128) → LayerNorm → GELU → Dropout → Linear(128→128) → LayerNorm → GELU → Dropout → Linear(128→64) → LayerNorm

• Stage 1 — GazeBase pretraining: regression on continuous fatigue score (881 subjects, 12,334 recordings, 1000 Hz)
• Stage 2 — SEED-VIG LOSO: binary alert/drowsy classification (12 subjects, 4,566 EEG windows, 200 Hz)

Features (16 clean oculomotor biomarkers)

Group	Features
Blink	blink_rate_pm, blink_count, mean_blink_ms
Fixation	num_fixations, mean_fix_ms, std_fix_ms
Saccade	num_saccades, mean_sac_ms, mean_sac_vel, peak_sac_vel, vel_std
Pupil	mean_pupil, std_pupil, pupil_range
Gaze dispersion	gaze_x_std, gaze_y_std

Evaluation

Metric	All subjects (n=12)	Excl. S6 outlier (n=11)
AUC-ROC	0.866 ± 0.137	0.902
F1-macro	0.833 ± 0.110	—
GazeBase R²	-0.225	—

Validated with Leave-One-Subject-Out cross-validation (12 folds).

Usage

import torch
from huggingface_hub import hf_hub_download

ck_path = hf_hub_download(repo_id="tdnathmlenthusiast/cognitive-fatigue-detector", filename="gaze_clean_checkpoint.pt")
ck = torch.load(ck_path, map_location="cpu")

model = GazeModel(ck["model_config"]["in_dim"])
model.load_state_dict(ck["model_state_dict"])
model.eval()

Datasets

• GazeBase — eye-tracking, 1000 Hz
• SEED-VIG — EEG vigilance, 200 Hz