Predictive Demand Algorithm

Tiered Forecasting Specification for Uganda's Health Supply Chain

---

Introduction

This document specifies a pragmatic forecasting framework for Uganda's health supply chain, informed by the 2025 Baseline Assessment conducted by IFRAD. The assessment revealed that 89% of facilities experience unreliable connectivity, widespread system fragmentation persists, and infrastructure constraints are primary drivers of stockouts.

Key Finding: Storage capacity demonstrated a strong negative correlation with stockout frequency (r = -0.695), confirming that physical infrastructure limitations must be treated as hard constraints rather than variables in any forecasting system.

This framework abandons a one-size-fits-all complex AI model in favor of a Tiered Forecasting Approach ensuring forecasting remains functional, understandable, and actionable at every level of the health system.

---

Framework Overview

Core Principles from Baseline Evidence

Principle	Implementation
Infrastructure as Binding Constraint	Storage capacity hard-coded as a limit in all algorithms, not an input variable
Connectivity is Intermittent	Forecasting must function primarily offline (89% unreliable connectivity)
Human Oversight is Critical	Overrides are not failures—they represent contextual knowledge algorithms cannot capture

---

Tiered Forecasting Architecture

┌────────────────────────────────────────────────────────────────────┐
│                         TIER 3                                      │
│              Machine Learning Forecasting                           │
│         (District/Central Level - Random Forest/XGBoost)            │
│                                                                     │
│   Purpose: Stockout risk prediction, district planning              │
└───────────────────────────────┬────────────────────────────────────┘
                                │
                                ▼
┌────────────────────────────────────────────────────────────────────┐
│                         TIER 2                                      │
│         Hierarchical Exponential Smoothing (HES)                    │
│              (District Server - Weekly Generation)                  │
│                                                                     │
│   Purpose: Second opinion for facility comparison                   │
└───────────────────────────────┬────────────────────────────────────┘
                                │
                                ▼
┌────────────────────────────────────────────────────────────────────┐
│                         TIER 1                                      │
│              Rule-Based Forecasting (Offline)                       │
│                 (Mobile App - Always Available)                     │
│                                                                     │
│   Purpose: Primary operational forecasting                          │
└────────────────────────────────────────────────────────────────────┘

---

Tier 1: Rule-Based Forecasting

Deployment: Embedded within the offline-first mobile application Connectivity: None required Target: All facilities, especially those with zero connectivity

Core Algorithm Logic

Base Calculation

Simple Moving Average of the past 3 months of consumption data.

Seasonal Adjustment

Pre-loaded fixed multipliers derived from multi-year DHIS2 national consumption patterns:

Commodity	Season	Adjustment
Antimalarials	April-June, October-December (rainy seasons)	+30%
Oral Rehydration Salts	Dry season months	+25%
Anti-snake venom	Planting and harvest seasons	+40%

Infrastructure Constraint Application

The final recommended order quantity is capped based on storage capacity:

Storage Rating	Cap Applied
Very Inadequate	50% of calculated demand
Inadequate	70% of calculated demand
Adequate	No cap

Rationale: This directly addresses the baseline finding that facilities with inadequate storage capacity experienced the highest stockout frequency (r = -0.695). The algorithm prevents recommendations for quantities that cannot physically be stored.

Cold Start Protocol

For new facilities or those with insufficient historical data (<3 months), the system matches against similar facilities using:

1. Facility Level: HC II, HC III, HC IV, or Hospital

2. Geographic Region: Karamoja, West Nile, Central, Eastern, etc.

3. Patient Volume Bracket: Low (<500/month), Medium (500-1500/month), High (>1500/month)

4. Commodity Class: HIV facilities vs. non-HIV, etc.

The system calculates median consumption from matched facilities, then applies seasonal adjustments and storage constraints.

Example: A new HC III in Karamoja with 800 monthly patients would use the median consumption from other HC III facilities in Karamoja serving 600-1000 patients per month.

Data Synchronization Protocol

Upload to District Server:

• Final approved order quantities

• Override logs (quantity changed, reason selected, user ID)

• Current stock levels

• Storage capacity status

Download from District Server:

• Updated regional baseline data (refreshed median consumption values)

• Revised seasonal multipliers

• Tier 2 forecasts for comparison

• Alerts or directives from district supply officers

---

Tier 2: Hierarchical Exponential Smoothing (HES)

Deployment: District-level servers Schedule: Weekly generation, pushed to facilities during sync Target: Facilities with intermittent connectivity

Model Selection

Hierarchical Exponential Smoothing (HES) using bottom-up reconciliation was selected for:

Advantage	Description
Scalability	Single model handles all facilities vs. 200+ separate ETS models
Resource Efficiency	80% reduction in computational overhead
Coherence	Bottom-up reconciliation resolves inconsistencies
Open Source	Available via Python (hierarchicalforecast) and R (hts, fable)
Interpretability	Clear hierarchical structure aids government stakeholder explanation

Hierarchical Structure

District Total
├── HC II Group (sum of all HC II facilities)
├── HC III Group (sum of all HC III facilities)
├── HC IV Group (sum of all HC IV facilities)
└── Hospital Group (sum of hospital facilities)

Data Sources

• DHIS2: Historical consumption patterns (12-24 months)

• e-LMIS: Stock level history and delivery schedules

• Tier 1 Uploads: Actual facility orders and consumption

• eAFYA: Patient service volumes (demand proxy)

Output and Use Case

Important: Tier 2 forecasts are not orders. They are a second opinion for facility staff to compare against their Tier 1 calculation.

User Interface Display:

┌──────────────────────────────────────────────────┐
│ ACT Tablets Forecast Comparison                   │
├──────────────────────────────────────────────────┤
│ Your Facility Forecast (Tier 1):    650 tablets  │
│ District Data Forecast (Tier 2):    720 tablets  │
│ Difference:                   +70 tablets (+10.8%)│
│                                                   │
│ Explanation: District data shows increased       │
│ malaria cases in nearby facilities this month.   │
└──────────────────────────────────────────────────┘

Purpose: Tier 2 identifies trends not visible from single-facility data:

• Emerging disease patterns across the district

• Seasonal patterns not yet reflected in 3-month windows

• Supply chain disruptions affecting multiple facilities

---

Tier 3: Machine Learning Forecasting

Deployment: Central cloud or district servers with reliable power/connectivity Target: District supply officers (not deployed to facilities)

Core Algorithm

Random Forest or XGBoost, selected for:

• Robust handling of missing data

• Ability to capture non-linear relationships

• Feature importance transparency

Feature Set

From Baseline Data:

• Storage capacity rating

• Average delivery lead time (days)

• Facility type and level

• Inventory count frequency

From National Systems:

• Budget utilization rates

• Historical stockout frequency

• Supplier delivery performance

• Disease surveillance alerts (DHIS2)

Derived Features:

• Days since last stockout

• Consumption velocity (units per day)

• Seasonality indicators

• Geographic risk factors (road accessibility, distance from hub)

Purpose and Output

Note: Tier 3 does not generate facility-level orders. Its outputs serve district supply officers for planning.

Use Cases:

• District-level demand forecasting for bulk procurement

• Stockout risk prediction (>80% probability in next 2 weeks)

• Proactive redistribution suggestions

Example Output:

HIGH RISK FACILITIES (Next 14 days)
───────────────────────────────────────────────────
Moroto HC III: 94% stockout risk for ACT
  → Recommendation: Emergency redistribution from
    Moroto Hospital (current stock: 180% of monthly need)

Abim HC II: 87% stockout risk for Amoxicillin
  → Recommendation: Expedite pending delivery
    (currently 3 days overdue)

---

Algorithm Workflow with Human-in-the-Loop

Facility-Level Workflow (Offline)

Step 1: Automated Forecast Generation

User opens the mobile app. The system automatically runs the Tier 1 Rule-Based Forecast for all tracked commodities.

Step 2: Forecast Review During Inventory

During weekly physical inventory count, the app displays:

┌────────────────────────────────────────────────────────┐
│ COARTEM (Artemether-Lumefantrine) 20mg Tablets         │
├────────────────────────────────────────────────────────┤
│ Current Stock:        420 tablets                      │
│ Recommended Order:    700 tablets                      │
│                                                        │
│ Calculation:                                           │
│   • 3-month average consumption: 800 tablets/month     │
│   • Malaria season adjustment:   +200 tablets (+25%)   │
│   • Calculated need:             1,000 tablets         │
│   • Storage limit applied:       700 tablets           │
│     (Inadequate storage = 70% cap)                     │
│                                                        │
│ [Accept]   [Adjust Quantity]   [Mark as Override]      │
└────────────────────────────────────────────────────────┘

Step 3: Human Override Capability

If user selects "Adjust Quantity":

• Enter new quantity

• Select reason from dropdown:

  • Disease outbreak

  • Delivery delay expected

  • Storage issue resolved

  • Expiry concern

  • Community health campaign

  • Other (free text)

Override is logged with timestamp and user ID.

Step 4: Local Order Approval

Facility in-charge reviews and approves. Order saved locally in SQLite database.

District-Level Workflow (Online)

Step	Action
1. Model Execution	District server runs Tier 3 ML model weekly
2. Officer Review	Supply officers review high-risk facilities, plan redistribution
3. Intervention	Send alerts, authorize redistribution, update delivery schedules

---

Model Evaluation and Validation

Primary Success Metric: Stockout Frequency and Duration

The ultimate measure is improved supply availability, not forecast accuracy:

• Percentage of facilities experiencing stockouts per month

• Average stockout duration (days)

• Commodity availability rate

Forecast Accuracy Metrics (Secondary)

Performance Level	MAPE Range	Action
Target	<20%	Routine monitoring
Acceptable	20-30%	Monitor closely
Review	>30%	Algorithm investigation triggered

Override Analysis

High override rates are signals, not failures:

• >50% override rate: Investigate forecasting assumptions

• Clustered overrides by reason: Systemic supply chain issue

• <5% override rate: Users may not be engaging with system

---

Technology Stack (Open Source)

Component	Technology
Tier 1 Mobile App	React Native, SQLite
Tier 2 Forecasting	Python, hierarchicalforecast library, pandas
Tier 3 ML	Python, scikit-learn (Random Forest), XGBoost
Data Sync	RESTful APIs over HTTPS with TLS 1.2+
Database	PostgreSQL (district/central), SQLite (mobile)

---

Transparency and Explainability

Every forecast must be explainable in simple terms:

Tier 1 Example:

"Based on your last 3 months of usage, adjusted for malaria season (+30%), limited by your storage capacity."

Tier 3 Example:

"High stockout risk due to: delivery delays (40% contribution), low recent stock levels (35%), nearby facility outbreaks (25%)."

---

Risks and Mitigation

Risk	Mitigation
Model Drift	Monthly automated retraining, quarterly governance review, automated MAPE alerts
Data Quality Issues	Quality dashboards, validation rules in app, facility feedback loop
User Trust	Transparent explanations, override capability, participatory design
Infrastructure Failure	Tier 1 fully offline, district server redundancy, CSV export fallback
Misuse/Gaming	Audit trails, anomaly detection, supervisory dashboards

---

Success Criteria

After 12 months of deployment:

Metric	Target
Stockout Reduction	30% reduction vs. baseline
Duration Improvement	40% reduction in average stockout duration
User Adoption	>80% of facilities actively using forecasts
Override Appropriateness	<20% of overrides flagged as anomalous
System Reliability	>95% uptime for Tier 1 offline functionality
Data Quality	>70% of facility data passing quality thresholds

---

Document Information

Field	Value
Version	2.0 (Post-Technical Validation)
Date	November 2025
Authors	Gideon Abako, Timothy Kavuma (IFRAD)
Validator	Ojok Ivan, Kyambogo University
License	CC BY 4.0

Key Changes in v2.0

• Tier 2 Model: Replaced standard ETS with Hierarchical Exponential Smoothing (HES)

• Cold Start Protocol: Added detailed facility matching criteria

• Data Sync Protocol: Specified bidirectional workflow

• Model Selection Rationale: Added comprehensive comparison section

• Open Source Implementation: Specified Python libraries for replication