neurogaze-gaze-lr-v1

Neurogaze Gaze LR v1

This repository contains the exact frozen model bundle used by the Neurogaze semifinal web application's deterministic legacy replay path. It is a custom JSON export of a logistic-regression classifier over 13 scanpath-geometry features. The scaler, classifier coefficients, Platt adjustment, and demo threshold are all contained in model.json.

Status

Research demonstration only. Not a medical device and not clinically validated. The current live tablet-camera flow deliberately does not use this classifier because its measurement contract does not match the training data. It must not be used to diagnose, reassure, or refer an individual child.

Intended use

• Reproduce the Neurogaze legacy replay demonstration.
• Audit feature extraction, browser/Python parity, and failure handling.
• Reproduce the proof-of-principle analysis on school-age eye-tracker data.

Prohibited use

• ASD diagnosis or clinical decision-making.
• Scoring children aged 16-30 months.
• Scoring live webcam/tablet sessions.
• Treating the exported demo threshold as clinically validated.

Model

• Type: logistic regression with standard scaling and Platt-on-logit adjustment.
• Input: one ordered vector of 13 scanpath-geometry features.
• Output: adjusted probability plus a demo-only operating point.
• Model version: neurogaze-gaze-lr-1.0.0.
• Exact model SHA-256: faf020c437131652272db72e21fb0deff18abb151b9089e6c0121cfb928dbcc8.

The canonical order is stored in model.json and feature_schema_v2.json:

1. ink_frac
2. centroid_x
3. centroid_y
4. std_x
5. std_y
6. span_x
7. span_y
8. radial_mean
9. radial_std
10. grid_entropy
11. n_active_cells
12. bbox_fill
13. aspect_ratio

See inference_example.py for a dependency-free reference implementation.

Training data and evaluation

The proof-of-principle analysis used 547 rendered eye-tracking scanpath PNGs from 54 school-age participants in the Carette dataset. Every recording from a participant was kept in the same fold, and primary metrics were computed after aggregation at participant level.

• Participant ROC-AUC: 0.8228.
• Participant-bootstrap 95% CI: 0.6974-0.9290.
• Record ROC-AUC: 0.7640.
• Demo threshold: 0.2242501759.
• Observed sensitivity/specificity at that internal operating point: 0.9231 / 0.1786.

The specificity is too low for clinical deployment. The threshold is retained only so the deterministic semifinal replay remains reproducible.

Domain limitations

• Participants averaged 7.88 years; the product target is 16-30 months.
• Source acquisition used an SMI Red-m eye tracker at 250 Hz.
• The tablet application estimates gaze from a consumer camera at roughly 20-30 fps.
• Rendered training PNGs do not preserve raw gaze timestamps.
• The model therefore validates a scanpath-geometry signal, not clinical performance in Indonesian toddlers or Posyandu deployment.

Bundle contents

• model.json: exact custom model weights and inference contract.
• ood_reference.json: frozen legacy out-of-distribution reference.
• participant_reference.json: participant-level explanation reference.
• feature_schema_v2.json: feature names, units, and bounds.
• training.json: full frozen training and evaluation output.
• degradation.json: tablet-proxy degradation study.
• dataset_manifest_carette_v2.json: provenance and redistribution warning.
• vision_model_card.json: declaration of the separate MediaPipe landmark dependency used for localization, not ASD classification.
• checksums.sha256: integrity hashes.

Dataset availability

Raw Carette images are intentionally not redistributed here. The local dataset manifest records that redistribution rights must be verified from the original distribution. The static-face dataset and its audit CNN are also excluded: the dataset is quarantined because provenance, an explicit license, documented consent, demographics, clinical label definition, and stable participant IDs are unavailable.

Third-party vision dependency

The web application separately uses a pretrained Google MediaPipe Face Landmarker bundle to localize 478 landmarks, including iris centers 468 and 473. It is not an ASD classifier. The exact runtime artifact has SHA-256 64184e229b263107bc2b804c6625db1341ff2bb731874b0bcc2fe6544e0bc9ff. The binary is not re-hosted in this repository because its redistribution terms must be established independently.

Source and reproducibility

• Source repository: https://github.com/raphaeladikara/Dataton_Project_Rizzz
• Product: https://dataton-project-rizzz.vercel.app

The source repository includes the feature extractor, training pipeline, browser implementation, parity fixtures, offline tests, grouped-split contracts, and model governance documentation.