XplainHealth — Uganda

Making AI health screening explainable and accessible

XplainHealth builds open-source explainable AI tools for community health in Africa. Our first product identifies children at risk of stunting in Uganda — giving community health workers a practical screening tool and district health officers real-time intelligence.

Try the Tool → See How It Works
Uganda DHS Data · Open Source · Universitat Pompeu Fabra · 2 Papers Under Review
Model Performance
0.704
AUC-ROC — Model with sibling-aware cross-validation on Uganda DHS data (4,530 children)
Both tools live now
Stunting Recall at Threshold
77.4%
Model correctly identifies 3 in 4 stunted children — threshold optimised for community screening
Children in Training Data
4,530
693 clusters · 3,194 mothers · All four Uganda regions · Zero sibling leakage
4,530
Children in training dataset — Uganda DHS
0.704
AUC-ROC with sibling-aware cross-validation
77.4%
Stunted children correctly identified by model
2
Live deployed open-source tools
2
Papers under review

1 in 4 Ugandan children is stunted. Most go undetected until it is too late.

Stunting — being too short for one’s age due to chronic undernutrition — affects 26% of Ugandan children under five. Its consequences are irreversible after age two: impaired brain development, lower school performance, reduced lifetime income, and a cycle that repeats through the next generation.

The problem is not that stunting cannot be prevented. The problem is that at-risk children are not identified early enough for intervention to make a difference. Uganda’s 170,000 community health workers have no practical tool for stunting risk screening.

26%
Uganda stunting prevalence (UDHS 2022)
3.6×
Higher risk for children of uneducated mothers
20%
Uganda target: reduce stunting to 20% by 2030 — not on track
170K
Community health workers with no stunting screening tool
⏳

The window closes at 24 months

Stunting in the first 1,000 days is largely irreversible. Current systems miss children until they reach a health facility — often years too late for effective intervention.

📏

MUAC cannot detect stunting

The only community screening tool — the MUAC tape — measures acute wasting, not chronic stunting. Health workers have no way to flag stunting risk during home visits.

🚫

Black-box AI cannot be trusted

Existing AI health tools give predictions without explanations. Health workers and policy makers cannot act on results they do not understand. Explainability is not optional — it is a clinical necessity.

📡

No real-time data for District Health Officers

DHOs allocate resources based on historical averages, not live data. Northern Uganda communities with 40%+ stunting receive no more support than lower-burden areas.

Our Solution

A two-tier explainable AI system — from community to district

XplainHealth connects community-level screening by health workers to real-time district-level monitoring — all powered by open-source explainable AI that every stakeholder can understand and trust.

📱
Community Screening App
Six-input mobile Progressive Web App for community health workers. Runs completely offline after first load — no internet needed for predictions. The AI model executes entirely in the browser. MUAC acts as a clinical safety override for severe acute malnutrition. Installs on any Android phone from the browser in one tap.
Offline PWA 6 inputs Android MUAC override Open source
nuru20.github.io/community-malnutrition-app →
Live
🔬
Research & Clinical Dashboard
Full 19-input stunting risk predictor with SHAP explainability — showing exactly which factors drive each prediction in plain language. Built for researchers, clinicians, and District Health Officers. Population tab shows live screening data from across the district in real time.
19 inputs SHAP XAI Population tab Streamlit Python
malnutrition-tool.streamlit.app →
Live
🗄
Shared Live Database
Both tools write every screening to a single PostgreSQL database. District Health Officers see all community and clinical screenings in real time — enabling targeted resource allocation based on live risk patterns rather than historical averages.
Supabase PostgreSQL Real-time District monitoring
github.com/Nuru20 — open source →
Live
How It Works

From community home visit to district intelligence

1
🏘
Health worker visits child
Community health worker opens XplainHealth on their phone during a routine home visit. Enters 6 simple inputs — child age, sex, MUAC, whether mother can read, children under 5, distance to clinic.
2
🤖
AI predicts and explains
The model runs instantly in the browser — no internet needed. Shows GREEN, YELLOW, or RED with a plain-language explanation of the top risk factors driving the result.
3
☁
Screening saved
When internet is available the record syncs to the shared database — district, worker code, risk level, and probability all stored, building a live picture of risk across the community.
4
📊
DHO acts on data
The District Health Officer opens the Research Dashboard. Sees real-time risk patterns across their district — which subcounties need urgent attention and where to deploy resources.
Evidence & Traction

Built on real data. Validated against national figures.

Our model is trained on the Uganda Demographic and Health Survey data — nationally representative data across all four regions. Prevalence estimates validated against published UDHS national figures.

✓ What We Have Built
✓
Community App live — offline PWA, works in rural Uganda without internet
✓
Research Dashboard live — SHAP explainability for every individual prediction
✓
Shared live database — both tools writing and reading in real time
✓
Both tools fully open source on GitHub
✓
Model validated against UDHS published national prevalence figures
✓
Novel sibling-aware GroupShuffleSplit cross-validation — eliminates data leakage in DHS prediction
✓
Interaction features improve AUC by +0.019 over baseline
✓
2 papers published
Team

Built by researchers who understand both the AI and the field

N
Nabuuso Nuru
Founder & Doctoral Researcher
Computer Science PhD researcher specialising in explainable AI. Designed and built both screening tools, the XGBoost model, the SHAP explainability pipeline, and the shared database architecture. Ugandan — with direct field access and deep understanding of the communities the tools serve.
Universitat Pompeu Fabra, Barcelona
V
Vladimir Estivill
Research Supervisor & Co-Author
Professor of Computer Science at Universitat Pompeu Fabra. Provides methodological supervision, research strategy guidance, and academic co-authorship. Deep expertise in machine learning and pattern recognition applied to complex real-world problems.
Universitat Pompeu Fabra, Barcelona
+
Uganda Partners
In Development
Actively building partnerships with Makerere University School of Public Health, Uganda Ministry of Health nutrition division, and District Health Officers for field validation, pilot study execution, and long-term deployment support.
Makerere University SPH — in progress

Partner with us to reach every child in Uganda

XplainHealth is seeking research collaborators, institutional partners in Uganda, and funding to scale our tools to district level across three Uganda regions. If you work in global health, nutrition, AI for social good, or international development — we want to hear from you.

🤝 We are open to

Research partnerships · Uganda MoH and NGO collaborations · Grant co-applicants · Pilot study partners · UNICEF and WHO conversations · Conference invitations · Investor meetings · Speaking opportunities

✉
Email the Foundernuru.nabuuso@upf.edu
🔬
Research Dashboardmalnutrition-tool.streamlit.app
📱
Community Screening Appnuru20.github.io/community-malnutrition-app
⚡
Source Code — GitHubgithub.com/Nuru20 — fully open source
👤
Researcher Profilehttps://nuru20.github.io/my_profile/