Datathon Teams & Tasks

35 participants • 7 teams

Applications of Data Science in Health AI using Vector Embeddings and Clinical Metadata

This datathon uses retinal image vector embeddings extracted from MedSigLIP, DINOv2, and RETFound models, combined with patient metadata to investigate clinically meaningful machine learning applications in diabetic retinopathy screening.

Teams & Assignments

Team 1

Group 1 — Multimodal Fusion Study

Members

Roland Mfondoum
Samuel Kyobe
EFFAH JEMIMA N'DA
Joel Sabiti
Hafeez Akinniyi

Goal

Study the impact of combining metadata and embeddings.

Suggested tasks

Compare metadata-only vs embeddings-only vs fused models
Experiment with fusion strategies
Perform ablation studies
Analyze the added value of metadata

Expected deliverables

Fusion performance comparison
Ablation analysis
Recommendations on multimodal learning

Team 2

Group 2 — Embedding Space Exploration

Members

Oluwole Olajide
David Nkurunziza
Nondumiso Mthiyane
Busisiwe Mlotshwa
Wisdom A Akurugu

Goal

Explore and visualize embedding representations learned by the models.

Suggested tasks

Perform PCA, t-SNE, or UMAP
Visualize clusters of disease severity
Compare MedSigLIP, DINOv2, and RETFound embeddings
Investigate separability of classes

Expected deliverables

2D embedding visualizations
Cluster analysis
Embedding comparison insights

Team 3

Group 3 — Risk Stratification & Screening Prioritization

Members

Keph Makoyi
Judith Mangeni
Norbert Nawe
Penina Safari
Tumai Muzorewa

Goal

Develop risk scoring systems for prioritizing diabetic retinopathy screening.

Suggested tasks

Build risk prediction models
Optimize sensitivity-focused thresholds
Investigate triage strategies
Discuss operational impact

Expected deliverables

Risk stratification framework
Screening prioritization strategy
Operational recommendations

Team 4

Group 4 — Embedding Quality Comparison Benchmark

Members

Adewemimo Olaosebikan
Natasha Lalloo
Anotida Marrian Chapunza
Nhlamulo Khoza
Imonikhe Kio

Goal

Perform a systematic benchmark comparison of MedSigLIP, DINOv2, and RETFound embeddings.

Suggested tasks

Train identical downstream models on each embedding type
Compare accuracy
Compare sensitivity
Compare specificity
Compare inference speed
Compare robustness
Analyze dimensionality and redundancy
Study how embeddings combine with metadata

Expected deliverables

Benchmark leaderboard
Comparative visualizations
Recommendations for healthcare deployment

Team 5

Group 5 — Explainable Health AI

Members

Eyad Elbahtety
Ezinne Uvere
Chika Ezeador
Oluwaseun Williams
OLUSOLA OGUNSANYA

Goal

Develop interpretable diabetic retinopathy prediction systems.

Suggested tasks

Train interpretable ML models
Perform SHAP analysis
Generate feature importance rankings
Interpret clinically meaningful variables

Expected deliverables

Explainability visualizations
Feature importance analysis
Clinical insights report

Team 6

Group 6 — Baseline DR Severity Prediction

Members

Emmy MUGISHA
Maria Gloria Nassuuna
Hillary Koros
Chantale Ruto
Brigitte UMUTONI

Goal

Predict diabetic retinopathy severity using embeddings only, metadata only, and combined multimodal features.

Suggested tasks

Build baseline ML models
Compare embeddings vs metadata vs combined features
Generate confusion matrices and evaluation metrics
Interpret model performance

Expected deliverables

Performance comparison table
Best-performing baseline model
Clinical interpretation of findings

Team 7

Group 7 — Bias & Fairness Analysis

Members

Celestin Twizere
Tendayi Mutangadura
Henriette Dukuzimana
Eric Katagirya
Jacqueline Wahura

Goal

Investigate whether model performance varies across patient subgroups.

Suggested tasks

Evaluate performance across sex and age groups
Investigate site or camera-related bias
Perform subgroup-specific error analysis
Discuss fairness implications

Expected deliverables

Fairness analysis report
Bias visualizations
Recommendations for equitable deployment

Challenge Task Groups

Goal: Study the impact of combining metadata and embeddings.

Suggested Tasks

Compare metadata-only vs embeddings-only vs fused models
Experiment with fusion strategies
Perform ablation studies
Analyze the added value of metadata

Expected Deliverables

Fusion performance comparison
Ablation analysis
Recommendations on multimodal learning

Goal: Explore and visualize embedding representations learned by the models.

Suggested Tasks

Perform PCA, t-SNE, or UMAP
Visualize clusters of disease severity
Compare MedSigLIP, DINOv2, and RETFound embeddings
Investigate separability of classes

Expected Deliverables

2D embedding visualizations
Cluster analysis
Embedding comparison insights

Goal: Develop risk scoring systems for prioritizing diabetic retinopathy screening.

Suggested Tasks

Build risk prediction models
Optimize sensitivity-focused thresholds
Investigate triage strategies
Discuss operational impact

Expected Deliverables

Risk stratification framework
Screening prioritization strategy
Operational recommendations

Goal: Perform a systematic benchmark comparison of MedSigLIP, DINOv2, and RETFound embeddings.

Suggested Tasks

Train identical downstream models on each embedding type
Compare accuracy
Compare sensitivity
Compare specificity
Compare inference speed
Compare robustness
Analyze dimensionality and redundancy
Study how embeddings combine with metadata

Expected Deliverables

Benchmark leaderboard
Comparative visualizations
Recommendations for healthcare deployment

Goal: Develop interpretable diabetic retinopathy prediction systems.

Suggested Tasks

Train interpretable ML models
Perform SHAP analysis
Generate feature importance rankings
Interpret clinically meaningful variables

Expected Deliverables

Explainability visualizations
Feature importance analysis
Clinical insights report

Goal: Predict diabetic retinopathy severity using embeddings only, metadata only, and combined multimodal features.

Suggested Tasks

Build baseline ML models
Compare embeddings vs metadata vs combined features
Generate confusion matrices and evaluation metrics
Interpret model performance

Expected Deliverables

Performance comparison table
Best-performing baseline model
Clinical interpretation of findings

Goal: Investigate whether model performance varies across patient subgroups.

Suggested Tasks

Evaluate performance across sex and age groups
Investigate site or camera-related bias
Perform subgroup-specific error analysis
Discuss fairness implications

Expected Deliverables

Fairness analysis report
Bias visualizations
Recommendations for equitable deployment

Suggested Evaluation Criteria

Criterion	Weight (%)
Problem Understanding	15
EDA & Data Cleaning	20
Technical Soundness	20
Innovation	15
Clinical Relevance	15
Deployment Feasibility	10
Presentation Quality	5

Teams are encouraged to focus on clinical relevance, explainability, fairness, and deployment feasibility in addition to model performance.

BRSET Data Dictionary

A Brazilian Multilabel Ophthalmological Dataset (BRSET) v1.0.1

PhysioNet — BRSET v1.0.1 • DOI

BRSET consists of 16,266 fundus images from 8,524 Brazilian patients. The labels.csv file contains identifiers, demographics, anatomical labels, diabetic retinopathy grading, quality parameters, and multi-label disease classifications for each image.

Dataset Files

File / Folder	Description
`fundus_photos`	16,266 color fundus photograph images (JPEG).
`labels.csv`	Table with image identifier, demographics, structural labels, diagnoses, and quality parameters.

Identifiers & Demographics

Column	Description
`image_id`	Image identifier.
`patient_id`	Patient identifier.
`camera`	Retinal camera (Canon CR-2 or Nikon NF505).
`patient_age`	Age of patient in years.
`patient_sex`	1 = male, 2 = female.
`exam_eye`	1 = right eye, 2 = left eye.
`nationality`	Patient's nationality.
`comorbidities`	Free text of self-reported clinical antecedents.
`diabetes`	Diabetes diagnosis.
`diabetes_time`	Self-reported time since diabetes diagnosis (years).
`insulin_use`	Self-reported insulin use (yes or no).

Anatomical Parameters

Column	Description
`optic_disc`	1 = normal, 2 = abnormal.
`vessels`	1 = normal, 2 = abnormal.
`macula`	1 = normal, 2 = abnormal.

Diabetic Retinopathy Classification

Column	Description
`DR_ICDR`	International Clinical Diabetic Retinopathy (ICDR) grade 0–4. 0 — No retinopathy 1 — Mild non-proliferative diabetic retinopathy 2 — Moderate non-proliferative diabetic retinopathy 3 — Severe non-proliferative diabetic retinopathy 4 — Proliferative diabetic retinopathy and post-laser status
`DR_SDRG`	Scottish Diabetic Retinopathy Grading (SDRG) grade 0–4. 0 — No retinopathy 1 — Mild background 2 — Moderate background 3 — Severe non-proliferative or pre-proliferative diabetic retinopathy 4 — Proliferative diabetic retinopathy and post-laser status

Quality Parameters

Column	Description
`focus`	1 = satisfactory, 2 = unsatisfactory.
`illumination`	1 = satisfactory, 2 = unsatisfactory.
`image_field`	1 = satisfactory, 2 = unsatisfactory.
`artifacts`	1 = satisfactory, 2 = unsatisfactory.

Classification Parameters (Multi-label)

Binary labels: 1 = present, 0 = absent.

Column	Description
`diabetic_retinopathy`	Diabetic retinopathy present.
`macular_edema`	Diabetic macular edema present.
`scar`	Scar (e.g. toxoplasmosis) present.
`nevus`	Nevus present.
`amd`	Age-related macular degeneration present.
`vascular_occlusion`	Vascular occlusion present.
`hypertensive_retinopathy`	Hypertensive retinopathy present.
`drusens`	Drusen present.
`hemorrhage`	Non-diabetic retinal hemorrhage present.
`retinal_detachment`	Retinal detachment present.
`myopic_fundus`	Myopic fundus present.
`increased_cup_disc`	Increased cup-to-disc ratio present.
`other`	Other pathology present.

Back to program