AI Framework for Humanitarian Health Supply Chain Optimization.
A technical and governance framework for offline first forecasting, stock visibility, and decision support.
A pragmatic framework for operating conditions on the ground.
This framework provides a comprehensive approach to optimising health supply chains in low resource humanitarian settings. Rather than a single complex AI system, it defines a distributed, tiered architecture that places core functionality on the mobile device at the facility level, where connectivity cannot be assumed.
It builds on earlier humanitarian innovation research conducted in Uganda with support from the UK Foreign, Commonwealth & Development Office and the Elrha Humanitarian Innovation Fund, and is maintained by Neuravox Foundation. The framework is the foundation phase of a broader humanitarian AI project now advancing into tool development.
Four commitments that shape every decision.
Offline first
All critical functions work without internet connectivity. Connectivity is treated as intermittent.
Human in the loop
AI provides recommendations; health workers make the final decisions. Overrides are part of the design.
Infrastructure aware
Storage capacity is a hard constraint, reflecting a strong negative correlation with stockouts.
Paper and digital harmony
The system augments existing paper workflows and generates outputs that match official registers such as HMIS 105.
Functionality placed where it is needed most.
The architecture distributes forecasting across three tiers so that the most critical functions never depend on connectivity, while districts and national teams add consolidation and oversight as connectivity allows.
- 01NationalCentral data hubAlways connected
- National data warehouse and long term trend analysis
- Integration with DHIS2, e-LMIS, eAFYA and the Nutrition Appointment Platform
- National dashboard: stockout trends, forecast accuracy, adoption
- 02DistrictTier 2 + Tier 3 forecastingPeriodic connectivity
- Tier 2 Hierarchical Exponential Smoothing for coherent forecasts
- Tier 3 machine learning (Random Forest / XGBoost) for stockout risk prediction
- District dashboard with redistribution planning and override audit logs
- 03FacilityTier 1 rule based (offline)No internet required
- Embedded rule based forecasting that runs on device in milliseconds
- Encrypted local SQLite storage and offline authentication
- Storage and budget aware recommendations approved by facility staff
Facility level offline decision support
An offline first Android application puts core functionality where it is needed most. Tier 1 rule based forecasting combines a 3 month moving average, seasonal multipliers, and storage constraints to produce a recommended order in under 100ms, fully offline. Recommendations are advisory: facility staff approve, adjust, or reject every order during the weekly inventory review, and the system generates paper outputs that match existing registers.
District level forecasting and consolidation
District services consolidate facility data when connectivity allows. Tier 2 Hierarchical Exponential Smoothing produces coherent forecasts across the facility hierarchy, while Tier 3 machine learning models flag facilities at high stockout risk and surface redistribution options from nearby surplus stock. District dashboards expose trends, forecast accuracy, and a full override audit trail.
National level analytics and oversight
A central data hub aggregates district data into a national warehouse, integrates with existing health information systems, and provides Ministry of Health teams with visibility into stockout trends, regional comparisons, forecast accuracy, and system adoption, supporting oversight without disrupting facility level decision authority.
Meeting facilities where they are.
A maturity model lets facilities integrate at a realistic level. Opportunistic sync is the primary target for the majority of facilities, with secure standards (TLS 1.3 in transit, AES-256 at rest, role based access).
Conflict resolution follows a simple rule: local human edits always win, with full audit trails on every change.
Paper centric
All records on paper registers. Focus: digitisation at source.
Single system
One digital system in use. Focus: standardising on the offline first app.
Opportunistic syncPrimary target
Primary target: offline first app syncs when connectivity allows.
Real time
Full API based exchange. Referral hospitals and district HQs only.
Codifying how decisions get made.
Each decision is defined by a clear protocol: trigger, system recommendation, decision authority, override process, data inputs, and expected outcome, so accountability is explicit. Tier 2 forecasts are advisory only, and local edits always win.
Use case library
Trust through verifiable records.
Paper remains the legal record; a digital versus paper variance over 5% triggers a mandatory physical audit within 14 days.
Automated alert thresholds for consumption spikes, stock variance, emergency order frequency, and expiry losses.
Immutable audit trail capturing user ID, timestamp, and pre/post values for every change, encrypted with AES-256.
DHIS2 morbidity cross checks so consumption spikes are validated against real disease data.
Seven principles for humanitarian AI.
The framework embeds humanitarian AI safeguards as design requirements, protecting communities, frontline workers, and decision authority.
- 01Do No Harm: recommendations stay advisory at every tier.
- 02Equity & Inclusion: refugee serving facilities weighted; storage aware caps protect the most constrained.
- 03Transparency & Explainability: plain language reasons for every recommendation.
- 04Accountability & Oversight: clear decision authority and quarterly algorithmic audits.
- 05Local Edits Always Win: human judgement overrides the model.
- 06Privacy & Data Protection: aligned with Uganda's Data Protection and Privacy Act (2019).
- 07Frontline Worker Protection: reduced administrative and cognitive burden.
Designed for measurable impact.
The framework includes a value for money methodology with worked ROI and social return formulas. The targets below are design objectives from the needs and baseline assessments; full quantification is planned for a later phase as facility financial data becomes available.
- Stockout duration
- Target 40% reduction
- Delivery delays (rainy season)
- Target 30% reduction
- Forecast accuracy (Tiers 2 to 3)
- ≥ 85% target
- Ordering cycle time
- From ~2 weeks to ~2 days
- Urban/rural accuracy disparity
- ≤ 5% target
The full specification on GitHub.
The complete technical, integration, operations, ethics, and assessment documentation is openly available in the Neuravox Foundation GitHub organization.