I.
Search strategy
Boolean search across MEDLINE, Embase, Web of Science, and Scopus, peer-reviewed using PRESS and scoped per-subspecialty with an AI/ML hedge.
§ Editorial Standards
How EyeModelHub
is built.
Every model card is the output of a defined seven-step process — pre-registered, dual-reviewed, quality-graded, version-controlled, and publicly auditable.
§01From Paper to Card
The seven-step pipeline.
One pipeline per subspecialty review · dual-reviewed at every gate.
II.
Title & abstract screening
Two reviewers screen each record independently in Covidence; discrepancies reconciled by a third reviewer.
III.
Full-text review
Eligible records are read in full and assessed against pre-registered inclusion criteria.
IV.
Data extraction
Two reviewers extract every Tier-1 and Tier-2 schema field per study, using controlled vocabularies for filterable categorisation.
V.
APPRAISE-AI scoring
Independent dual-reviewer scoring across 24 items in 6 subdomains; total reconciled and rendered as a quality grade.
VI.
Model-card publication
Cards are written in MDX, Zod-validated at build, dual-reviewed in pull requests, and deployed automatically on merge.
VII.
Living updates
Annual full re-search per subspecialty; continuous user submissions screened within four weeks. Versions follow semver.
§02Featured Model · Anatomy of a Card
Every model is a structured artefact.
Tier 1 for essence Tier 2 for evidence Tier 3 for provenance.
Model Name
First Author et al. · YYYY · Country · doi:xx.xxxx/xxxxxxx
Task · Task Anatomy · Anatomy A Anatomy · Anatomy B Modality · Modality Architecture · Architecture Maturity · Maturity Quality · Grade
Headline performance metric — value (95% CI x.xx–x.xx)
vs. comparator, brief context
"Plain-language summary of where the model is and isn't safe to use."
Caveats on training cohort, external validation, and subgroup reporting.
Total 69/100
Clinical relevance
3/4
Data quality
17/24
Methodological conduct
14/20
Robustness
14/20
Reporting quality
8/12
Reproducibility
13/20
- Reviewer
- Reviewer Name (ORCID)
- Appraised
- YYYY-MM-DD
- Card version
- 1.0.0
- Card status
- Current
- Code
- github · license
- Regulatory
- Status
- External val.
- N cohorts · protocol
- Demographics
- N=… · sex · ancestry
Validation note Brief note on validation strength or weakness, surfaced where it matters most.
§03In Detail
Methodology in depth.
The protocol, the rubric, the reviewer roster, the inclusion log.
01 Living scoping review process
+
Living scoping review process
EyeModelHub follows the PRISMA-ScR extension to scoping reviews. Each subspecialty review is a pre-registered, dual-reviewed mapping of the literature, refreshed annually with continuous interim updates from user submissions.
02 Inclusion criteria
+
Inclusion criteria
Peer-reviewed primary research describing AI/ML applied to a recognised ophthalmology subspecialty. Both research prototypes and regulator-cleared products are eligible if reported in primary literature. Excluded: preprints, reviews, editorials, conference abstracts, and non-human studies.
03 Model-card schema (v1.0)
+
Model-card schema (v1.0)
A 3-tier schema. Tier 1 (11 fields) is always visible. Tier 2 (~36 fields) is the expanded view. Tier 3 (~10 fields) is administrative metadata. Multi-value and enum fields use controlled vocabularies; "Not reported" is a legitimate value.
04 APPRAISE-AI grading
+
APPRAISE-AI grading
A 24-item, 6-subdomain critical-appraisal tool produces a 0–100 total per study. Subdomains: clinical relevance, data quality, methodological conduct, robustness, reporting quality, and reproducibility. Bands: Low <40, Moderate 40–59, High 60–79, Very High 80+.
05 Update cycle
+
Update cycle
Annual full re-search per subspecialty on the anniversary of the search closing date. Continuous user submissions reviewed within four weeks. Card versions follow semver — bump on any content change.
06 Reviewer adjudication
+
Reviewer adjudication
Every model card has at least two reviewers: one extracts, one adjudicates. APPRAISE-AI is scored independently by both. Discrepancies are resolved by discussion or by a third reviewer. Reviewer attribution is opt-in.
§04Living Document
∗
Methodology may evolve.
As the registry grows across subspecialties, methodological details — particularly around APPRAISE-AI scoring conventions, controlled vocabularies, and update cadence — may be refined. Material changes will be versioned, dated, and announced. The schema follows semantic versioning; this page is the living source of truth.