Synthetic blink data and functional-evaluation outputs of a low-cost automated blink analysis system for pediatric frequent blinking (proof-of-concept)

Because ethics approval for paediatric data collection had not yet been obtained, an end-to-end pipeline (face/eye landmark detection, EAR computation, blink and eye-rubbing event detection, parameter extraction, and four-class XGBoost classification) was exercised on synthetic data; the dataset contains no patient or personal information.

Contents: (1) synthetic blink-parameter data — training records (n = 320; four balanced classes: normal, dry eye disease, allergic conjunctivitis, tic disorder), a held-out test set (n = 80), 20 boundary-testing records (SIM-01-SIM-20, flagged as simulation data excluded from training), and the data generator script (interval specification in Table 2; independent uniform sampling; random_state = 42); (2) complete structured outputs of five USB-camera functional-evaluation sessions — per session blink_params.csv, ear.csv (per-frame timestamp, EAR, rubbing flag; 117-523 rows), and result.json (four-class probabilities and predicted class); (3) sensitivity analyses — detection-threshold variation over +/-20% of the defaults (Supplementary Table S3) and frame-rate decimation from 28-30 fps to 15/10/7.5 fps (Supplementary Table S4).