EyeAI combines DERIVA-ML with GPU-based notebooks to create a fully traceable pipeline for glaucoma detection from fundus photographs, spanning data ingest, feature extraction, model training, and clinical review. By treating each dataset, feature set, workflow, and execution as a first-class, versioned artefact, the project shows how DERIVA-ML delivers continuous FAIRness for real-world medical AI.
Why DERIVA-ML underpins EyeAI
| Challenge | DERIVA-ML Capability |
|---|---|
| Multimodal, noisy clinical data | Unified domain + ML schemas; data checked and versioned on ingest |
| Rapid feature & model iteration | Model-driven UI updates automatically when schema changes |
| Reproducible, auditable runs | Six-artefact data model (Dataset, Feature, Workflow, Execution, Asset, CV) captures every run |
| Sensitive patient data | Fine-grained Globus Auth roles (reader / writer / curator) with embargo support |
Impact to date¹
- >70 datasets, 130 executions, 142 trained models stored with DOIs
- 13 clinicians provided independent labels; inter-rater metrics computed automatically
- Label-efficiency study trained 142 models across 36 data subsets, with 71 retrains done by editing only execution configs
- Optic-nerve detector re-trained in one day after performance drift, demonstrating end-to-end reproducibility
Typical workflow
- Ingest — fundus JPEGs and metadata uploaded; Minid + BDBag generated.
- Feature engineering — VGG19 model crops optic-nerve region, stores SVG as Image Annotation Feature.
- Model training — VGG19 and RETFound classifiers trained via notebook; parameters, ROC figure, and model file auto-uploaded as Execution Assets.
- Review — Chaise UI lets ophthalmologists inspect predictions and trace back to raw images.
Lessons for other ML health projects
- Continuous FAIRness (“all data, all the time”) prevents data cascades.
- Controlled vocabularies plus Minid-based datasets make label auditing straightforward.
- Workflows stay in plain notebooks; DERIVA-ML adds the provenance layer without locking teams into new tools.
¹ Figures from Li Z. et al., “From Data to Decision: Data-Centric Infrastructure for Reproducible ML in Collaborative eScience,” arXiv 2506.16051 (2025) and Li Z. et al., “DERIVA-ML: Continuous FAIRness for Reproducible Machine Learning Models,” arXiv 2407.01608 (2024).