macular/diabetic-retinopathy-screening-benchmark-africa

DR-Africa-Benchmark — Screening-Prevalence-Corrected, Fairness-Instrumented DR Evaluation

An evaluation benchmark for diabetic-retinopathy grading under African screening conditions. It does not introduce new labels; it introduces evaluation validity — per-record importance weights that reweight a referral-skewed image set to real Sub-Saharan-Africa population prevalence, plus synthetic subgroup metadata for fairness reporting.

Version 1.0.0 · core dr_synth 1.0.0 · part of the DR-Africa dataset family (see also dr-grading and dr-progression).

---

Abstract

A DR model that scores well on a curated, referral-skewed test set can fail in real screening, where the overwhelming majority of imaged eyes have no retinopathy. In Sub-Saharan-Africa (SSA) population-based studies, any-DR prevalence is ~30–35% and proliferative DR ~1% [1, 10], whereas curated research sets (including the source images here) over-represent severe disease. Evaluating on the curated distribution therefore inflates apparent screening performance and hides subgroup disparities. DR-Africa-Benchmark corrects this at evaluation time. For each of 3,554 fundus records it provides (i) a screening_weight that reweights the empirical grade distribution to an SSA population-based target, and (ii) synthetic subgroup metadata (region, sex, age_band, urban, care_setting) for fairness slicing. We verify by audit that the weighted grade prevalence reproduces the target to within 0.001 and that every subgroup is represented. The benchmark is a held-out evaluation set for models trained on the sibling datasets. As with the whole family, the images are not African and the metadata is synthetic; the benchmark tests how a model holds up under a screening-prevalence setting, and is not itself African imaging.

---

1. Introduction

Two distributions matter for a screening DR model and they are very different: the training/research distribution (often enriched for disease) and the deployment distribution (screening, dominated by no-DR). Reporting a single accuracy number on the former is one of the most common ways a medical-imaging result fails to translate. The standard remedies are (a) importance-weighted evaluation toward the deployment prevalence, and (b) disaggregated, per-subgroup reporting to surface disparities that a pooled metric conceals.

DR-Africa-Benchmark provides both. It is not a training set; it is for measurement. By supplying weights and subgroup labels alongside the same images used elsewhere in the family, it lets any DR grader be re-scored as if it were running on an SSA screening population, and checked for fairness across demographic and access groups.

Contributions.

1. Per-record screening-prevalence importance weights, audited to reproduce an SSA population-based target.
2. Synthetic subgroup metadata enabling fairness/robustness evaluation.
3. A reproducible audit of weighted prevalence and subgroup representation.

---

2. Background and Related Work

Prevalence gap. SSA population-based DR prevalence (~30–35% any DR, ~1% PDR [1, 10]) versus the referral-skewed curated source distribution (PDR ~31% here) motivates prevalence correction.

Importance-weighted evaluation. Under label shift, an estimator of deployment performance reweights each test example by target(y)/source(y). We apply this at the grade level (a label-shift assumption: image appearance given grade is held fixed — see §8 for why this addresses prevalence but not appearance shift).

Disaggregated evaluation / fairness. Following model-card and subgroup- reporting practice, we attach strata (region, sex, age_band, urban/rural, care_setting) so that sensitivity/specificity can be reported per group rather than pooled.

---

3. Construction

3.1 Image source

APTOS 2019 (India) [11], 224×224 RGB, five DR-grade folders. Not African; transfer/evaluation source only (§8–§9).

3.2 Screening weights

Let obs(g) be the empirical proportion of grade g in the image set and target(g) the SSA population-based prevalence. Each record receives

screening_weight(g) = target(g) / obs(g),     then mean-normalised to 1.

The target prevalence (proportion of diabetics by grade) is

Grade	Target
No_DR	0.70
Mild	0.13
Moderate	0.10
Severe	0.05
Proliferate_DR	0.02

derived from SSA population-based estimates [1]. Weights are an evaluation device: they correct prevalence without fabricating images for rare-but-here grades. They must not be used to up-sample images you do not have.

3.3 Subgroup metadata

region, country, sex, age, age_band (<40, 40–54, 55–69, 70+), urban, and care_setting are generated by the shared dr_synth core (seed 42) from SSA-grounded parameters (dr_synth/params.yaml). They are synthetic (§8).

3.4 Splits

Stratified 70/15/15 on dr_grade (split seed 7), as a split column and data/{train,validation,test}.parquet. No image-id leakage (audit-verified). For pure benchmarking, the test split is the primary evaluation set; the others support weighted-model selection if desired.

---

4. Dataset Statistics

Field	Type	Description
image_id, image_path	string	fundus image reference
dr_grade, dr_grade_name	int/string	grade label (0–4)
screening_weight	float	importance weight for prevalence-corrected eval
region, country, sex, age, age_band, urban, care_setting	mixed	subgroup metadata (synthetic)
split	string	train/val/test

Grade distribution (n = 3,554): No_DR 968 · Mild 527 · Moderate 395 · Severe 575 · PDR 1,089 (referral-skewed). Splits: train 2,487 · validation 532 · test 535.

Subgroups (shares): region — west 30.2%, east 28.9%, southern 17.4%, central 12.2%, north 11.4%; sex ~52% F; age bands span <40 to 70+; urban ~55%; care setting clinic ~74% / screening ~26%.

---

5. Distribution Audit (Validation)

python build.py regenerates audit/AUDIT.md; verdict PASS. Key results:

5.1 Weighted prevalence reproduces the target

Grade	Weighted prevalence	Target	Status
No_DR	0.7000	0.70	pass
Mild	0.1300	0.13	pass
Moderate	0.1000	0.10	pass
Severe	0.0500	0.05	pass
Proliferate_DR	0.0200	0.02	pass

Maximum deviation < 0.001.

5.2 Subgroup representation

Every region, sex, age_band, and urban stratum exceeds a 2% minimum share, so per-subgroup metrics are estimable. Per-stratum label-proportion tables are written to audit/tables/subgroup_*.csv.

5.3 Integrity

No cross-split id leakage; age and screening_weight within range; no missing values.

---

6. How to Use the Benchmark

Compute weighted metrics (deployment-prevalence estimate) and per-subgroup metrics (fairness):

import pandas as pd, numpy as np
from datasets import load_dataset

test = load_dataset("macular/diabetic-retinopathy-screening-benchmark-africa", split="test").to_pandas()
preds = model_predict(test["image_path"])           # your grader

# prevalence-corrected accuracy
w = test["screening_weight"].to_numpy()
acc = np.average((preds == test["dr_grade"]).to_numpy(), weights=w)

# fairness: per-region sensitivity for sight-threatening DR (grade >= 3)
for region, g in test.groupby("region"):
    stdr = g["dr_grade"] >= 3
    sens = np.average((preds[g.index] >= 3)[stdr], weights=w[g.index][stdr])
    print(region, f"STDR sensitivity (weighted): {sens:.3f}")

Local: pip install -r requirements.txt && python build.py && python loader.py --split test. loader.py returns (image, grade, screening_weight, subgroup_dict).

---

7. Intended Uses

Intended. Held-out, prevalence-corrected evaluation of DR graders; fairness and robustness auditing across demographic/access strata; reporting that separates referral-set accuracy from screening-realistic accuracy.

Out of scope. Training (this is an evaluation resource); resampling rare images to "hit" the weights; clinical claims about real African patients.

---

8. Limitations

1. Prevalence shift, not appearance shift. Weights correct the grade distribution under a label-shift assumption. They do not correct image-appearance domain shift (African fundus pigmentation, smartphone-camera optics); a model can still fail on real African images even after weighting.
2. Non-African images. As above, the imagery is APTOS (India).
3. Synthetic subgroups. region, age, sex, etc. are generated, so fairness slices reflect the encoded distribution and are valid for methodology and tooling, not as evidence about real African subpopulations.
4. Single target prevalence. One SSA population-based target is used; regional prevalence varies and can be substituted by editing the target in params.yaml.

---

9. Ethical Considerations and Data Governance

• Subgroup labels measure, they do not certify: they let you measure disparities, but because they are synthetic they cannot prove real-world equity. Real subgroup data and local validation are needed for that.
• Synthetic flagged (synthetic = True); imagery provenance stated plainly.
• Data sovereignty: substituting real African imagery and real subgroup attributes must follow local consent, ethics approval, and data-sovereignty principles.
• Purpose: to make screening-realistic, disaggregated evaluation the default for DR models intended to serve African populations.

---

10. Licensing, Access, and Provenance

• Images: APTOS 2019-derived (Aravind Eye Hospital, India), released under the Apache License 2.0; redistribution permitted with attribution. Bundled images are provided under Apache-2.0; see LICENSE and ATTRIBUTION.md [11].
• Weights, metadata, code, audit: reproducible from dr_synth/params.yaml and fixed seeds; released under Apache-2.0.
• Provenance fields per row: synthetic, image_population, ehr_params_version.

---

11. Maintenance and Versioning

Regenerable via python build.py. To target a different regional prevalence, edit audit_targets.screening_prevalence in params.yaml and rebuild; the audit re-verifies the weighted distribution automatically. Core/version stamping as in the rest of the family.

---

12. References

1. Diabetic retinopathy in Sub-Saharan Africa: prevalence and regional variations — a systematic review and meta-analysis. BMC Ophthalmology, 2025. https://link.springer.com/article/10.1186/s12886-025-04589-5
2. Epidemiology of diabetic retinopathy and maculopathy in Africa: a systematic review. 2015. https://pmc.ncbi.nlm.nih.gov/articles/PMC4463765/
3. APTOS 2019 Blindness Detection. Kaggle. https://www.kaggle.com/competitions/aptos2019-blindness-detection
4. Gebru et al. Datasheets for Datasets. Communications of the ACM, 2021.

(Reference numbers are shared across the DR-Africa family; the full bibliography appears in dr-progression.)

---

Citation

@misc{drafrica_benchmark_2026,
  title        = {DR-Africa-Benchmark: Screening-Prevalence-Corrected,
                  Fairness-Instrumented Diabetic-Retinopathy Evaluation},
  author       = {DR-Africa dataset family},
  year         = {2026},
  note         = {Version 1.0.0. Images derived from APTOS 2019 (India);
                  weights and subgroup metadata are epidemiology-grounded
                  synthetic data.},
  howpublished = {Hugging Face Datasets}
}