Fundamentals of Data Science

Day 1 | WISE Workshop — Addis Ababa 2026

Sean Sylvia, Ph.D.

February 2, 2026

Table of contents

• Learning Objectives
• The Data Revolution in Health
• What Is Data Science?
• The Fundamental Question
• Ethiopia’s Pharmaceutical Supply Chain
• Three Paradigms of Quantitative Analysis
• The Prediction-Simulation Connection
• ML Is About Prediction
• What This Workshop Will Build
• The Data Science Mindset
• Your Role This Week
• Discussion: Your Supply Chain Challenges
• Key Takeaways
• What’s Next
• Questions?

Learning Objectives

By the End of This Session, You Will…

1. Distinguish between statistics, econometrics, and machine learning approaches
2. Identify when prediction vs. causal inference is the right goal
3. Understand how ML fits into the workshop’s overall pipeline
4. Articulate prediction challenges in your own supply chain context

This session is about framing — “why” and “when” rather than “how.” The technical “how” comes in Days 2-3. Connect this to their actual work throughout.

The Data Revolution in Health

A Transformative Moment

Health systems across Africa are experiencing an unprecedented data explosion.

What’s Changing

• Electronic inventory systems
• Digital health records
• Mobile health platforms
• Real-time reporting

The Opportunity

• Earlier detection of problems
• More efficient resource allocation
• Evidence-based policy decisions
• Improved health outcomes

Discussion: What data sources does your organization currently use? What challenges do you face?

This is an opening discussion moment — give 2-3 minutes for informal sharing. Note down examples they mention for later reference.

AI as Economic Transformation

Economics of Digitization:

Reduction in (Marginal) Costs

Cost Type	Healthcare Application
Search Costs	Instant medical literature access, symptom lookup
Replication Costs	Digital health records, AI models deployed at scale
Transportation Costs	Telemedicine, remote monitoring, digital consultations
Tracking Costs	Continuous patient surveillance, automated data collection
Verification Costs	Digital identity, online reputation systems, trust mechanisms

The Dream: AI-Enabled Healthcare

What AI Can Do

• Extend specialist expertise to frontline workers
• Standardize care quality across settings
• Enable continuous learning from practice
• Reduce diagnostic delays

The Sustainability Question

• Can AI systems improve and remain equitable over time?
• Will they work for hard-to-reach populations?
• How do we design for long-term rather than short-term gains?

The Leapfrog Opportunity {.center}

The Choice LMICs Face

Option 1: Follow the HIC Path

• Safe, proven trajectory
• Limited by legacy infrastructure
• Ceiling set by institutional debt

Option 2: Leapfrog to the Frontier

• Skip legacy systems entirely
• Higher potential productivity
• But significant uncertainty

The Question

How do we capture the leapfrog opportunity while managing the risk?

[~45 sec] Open with opportunity, not problems. LMICs have a unique chance to skip the mistakes of HIC health IT. Mobile penetration is high, CHW networks exist—the infrastructure is there. But leapfrogging comes with uncertainty. The question is: how do we design systems that capture this opportunity while managing the risk?

AI is Different

The Complexity

• Data shapes decisions → decisions change behaviors
• Behaviors generate new data → feedback loops
• Yesterday’s patterns ≠ tomorrow’s reality
• Unknown unknowns we can’t predict upfront

The Core Insight

Health systems are information-processing and decision-making factories.

Digital systems transform the factory floor—not just how care is delivered, but what care, when, and to whom.

We must learn about the system before deploying—and design to learn continuously after.

[~90 sec] Health systems are fundamentally information-processing machines. Digital health doesn’t just make processes faster—it changes what decisions get made, when, and for whom.

But here’s the complexity: these are sociotechnical systems with feedback loops. Data shapes decisions, decisions change behaviors, and those changed behaviors generate new data. The system you’re trying to improve keeps shifting under your feet.

This means unknown unknowns—emergent effects we can’t predict upfront. You can’t retrofit sustainability onto a digital health system after deployment. The sociotechnical has to be designed in from the start. We need to learn about the system before deploying, and design systems that continue learning after.

AI Often Fails In Healthcare

System	Scope	Key Failure	Real-World Impact
Epic Sepsis Model	56% of US hospitals	Missed 67% of cases, false alerts 88% of the time	Delayed treatment, alert fatigue
IBM Watson Oncology	Major hospitals globally	Unsafe treatment recommendations	$62M+ wasted, quietly discontinued
SkinVision App	Consumer market	235% worse performance with rotation	Missed melanomas, health disparities
COMPAS (Criminal Justice)	46 US states	44.9% vs 23.5% false positive rates	Systematic racial bias in sentencing

Pattern: Technical validation ≠ Real-world success

Each row represents millions in investment and thousands of lives affected. The pattern is clear: systems that pass technical validation fail when deployed in complex social systems. Include source citations in appendix.

Six Challenges to AI Adoption

ECAM: Four Principles for Sustainable AI

ENGAGE

Stakeholders First

• Map stakeholders before building
• CHWs, patients, officials
• Co-design, not consultation

CALIBRATE

Local Context

• Train on local populations
• Local languages, diseases
• Don’t import HIC models

AUDIT

Bias Continuously

• Equity audits by subgroup
• Gender, geography, SES
• Check for hallucination

MONITOR

Adapt Continuously

• Human override mechanisms
• Retraining triggers
• Performance dashboards

Templin et al. (2024) PLOS Digital Health | Templin et al. (2025) npj Digital Medicine

[~120 sec] Based on our work synthesizing challenges and evaluation frameworks for AI in healthcare, we propose four principles—ECAM.

Engage: Before you write a line of code, map your stakeholders. Community health workers, patients, local officials. Co-design, not just consultation.

Calibrate: Don’t import models trained on US or European populations. Calibrate to local languages, disease patterns, care pathways.

Audit: Build in equity audits from day one. Test across subgroups—gender, geography, socioeconomic status. Check for hallucination.

Monitor: Design for feedback. Human override mechanisms. Continuous retraining triggers. Performance dashboards.

What Is Data Science?

The Intersection of Three Fields

Data science sits at the intersection of domain knowledge, statistics, and computing

Your domain expertise is essential — algorithms alone cannot solve health system challenges.

Emphasize that participants bring the domain expertise. We’re providing methods, but their knowledge makes it meaningful.

The Fundamental Question

Can We Predict Health Outcomes Accurately?

This is the core question that motivates modern data science in health.

Traditional Statistics

• Focus on inference (\(\hat{\beta}\))
• Understanding relationships
• Hypothesis testing
• Causal interpretation

Machine Learning

• Focus on prediction (\(\hat{y}\))
• Minimizing forecast error
• Out-of-sample performance
• Pattern recognition

Both approaches are valuable — the key is knowing when to use each.

This is THE key distinction for the workshop. Beta-hat = we care about the coefficient (causation). Y-hat = we care about the prediction (forecasting).

Supply Chain Prediction Tasks

How can prediction improve pharmaceutical supply chains?

Task	Traditional Approach	ML Approach	Benefit
Demand Forecasting	ARIMA, exponential smoothing	Random forests, gradient boosting	Captures complex patterns
Stockout Risk	Threshold rules	Ensemble classification	Earlier warning
Lead Time Estimation	Average + buffer	ML with facility features	Facility-specific estimates
Expiry Prediction	Simple ratios	Pattern learning	Waste reduction

These are concrete examples they’ll work with in Days 2-3. Ask which of these resonate with their work.

Ethiopia’s Pharmaceutical Supply Chain

The Context

Scale & Complexity

• EPSA manages distribution nationwide
• 3,500+ health facilities served
• 1,000+ essential medicines tracked
• Multiple distribution tiers

Key Challenges

• Medication stockouts
• Demand variability
• Transportation constraints
• Information gaps

This is their context — make sure examples throughout connect back here. The WISE project is specifically about building Ethiopian capacity.

Three Paradigms of Quantitative Analysis

Statistics, Econometrics, and Machine Learning

Approach	Primary Goal	Key Question
Statistics	Estimation & inference	“What is the relationship?”
Econometrics	Causal identification	“What is the effect?”
Machine Learning	Prediction	“What will happen next?”

The “Beta-Hard” vs “Y-Hard” Distinction

• Beta-hard tasks: We care about why — causal or structural inference
• Y-hard tasks: We care about what — accurate prediction is the goal

This terminology comes from the economics/ML literature. Beta-hard = hard to estimate the causal parameter. Y-hard = hard to predict the outcome accurately.

When to Use Each Approach

Scenario	Best Approach	Example
Testing if an intervention works	Econometrics	Does the new distribution policy reduce stockouts?
Understanding which factors matter	Statistics	Which facility characteristics correlate with stockouts?
Forecasting next month’s demand	ML	How much of Drug X should we deliver to Region Y?
Optimizing resource allocation	ML	Which facilities are highest risk for stockouts?

In practice, we often combine approaches — this workshop shows you how.

The Prediction-Simulation Connection

How This Workshop Fits Together

The workshop pipeline: from data through ML models and simulation to policy insights

ML predictions feed simulation models — enabling “what if” policy analysis.

This is the workshop’s organizing framework. Data → ML → ABM → Policy. Each day builds toward this pipeline.

ML Is About Prediction

An Important Clarification

Important!

Machine Learning is about prediction. Period.

It does not magically solve the causal inference problem.

What ML CAN do:

1. Provide accurate forecasts for planning and allocation
2. Identify high-risk facilities or products for intervention
3. Help strengthen causal analyses in specific ways (we’ll see this later)

What ML CANNOT do:

• Tell you why something happens
• Prove that changing X will cause Y to change
• Replace careful study design for causal questions

This is a key takeaway to prevent misuse of ML. Many people think ML = magic causal machine. It’s not.

A Concrete Example: Poverty Targeting

The Task

Governments need to identify poor households for social assistance programs.

Traditional approach: Survey everyone (expensive, slow)

ML approach: Predict consumption from observable characteristics

How It Works

1. Survey a sample of households
2. Train ML model: characteristics → consumption
3. Apply model to rank all households
4. Target resources to highest-need predictions

Result: Same budget reaches more people in need

Key insight: This is purely predictive — the model doesn’t explain why households are poor, but it identifies who is poor.

This is a real-world example from development economics. Similar logic applies to supply chain: predict which facilities will stock out. We don’t need to know why to allocate resources effectively.

What This Workshop Will Build

Day-by-Day Preview

Day	Focus	What You’ll Learn
Day 1	Foundations	Reproducible research, project setup
Day 2	ML Pipeline	Forecasting models, evaluation methods
Day 3	Integration	Connecting ML outputs to simulation
Day 4	ABM Deep Dive	Agent-based modeling for policy

By the end: You’ll have built a complete pipeline from data to policy-relevant predictions.

The Data Science Mindset

Five Principles for This Week

1. Prediction is not causation — Always ask “what question am I answering?”

2. Out-of-sample performance matters — Models that memorize data are useless

3. Domain expertise is essential — Algorithms can’t replace your knowledge

4. Reproducibility is non-negotiable — Code and documentation matter

5. Perfect is the enemy of good — A useful model today beats a perfect model never

These five principles should guide their thinking throughout. Return to these when they face decisions about methods.

Your Role This Week

Bringing Your Expertise

What You Bring

• Understanding of your health system
• Knowledge of supply chain realities
• Awareness of data quality issues
• Ideas for relevant applications

What We’ll Provide

• Technical methods and tools
• Hands-on practice opportunities
• Templates and code examples
• Frameworks for thinking about problems

The best outcomes come from combining both — methods without context are just math.

Discussion: Your Supply Chain Challenges

Let’s Hear From You

In small groups (3-4 people), discuss:

1. What prediction problems do you face in your work?
   • What would you want to forecast more accurately?

2. What data exists that could help?
   • What’s available? What’s missing?

3. What decisions would better predictions improve?
   • How would you use the predictions?

Time: 5-7 minutes in groups, then brief sharing

Circulate during small group discussion. Take notes on challenges mentioned — these can inform examples later. Have 2-3 groups share briefly with full room.

Key Takeaways

Remember This

Prediction vs. Causation

What ML gives you:

• Accurate predictions (\(\hat{y}\))
• Pattern recognition
• Risk stratification

What ML does NOT give you:

• Causal effects (\(\hat{\beta}\))
• Mechanism understanding
• Policy counterfactuals

For “Why?” questions, we need careful study design. For “What’s next?” questions, ML is powerful.

What’s Next

Rest of Today’s Schedule

Time	Session	Description
10:30-10:50	Break
10:50-12:30	Data Science in Public Health Policy	Ozawa
12:30-1:30	Lunch
1:30-3:00	ML Foundations: Core Concepts	Sylvia
3:00-3:20	Break
3:20-5:00	Hands-on: Setup + ML Demo	Sylvia

Afternoon hands-on: We’ll set up everything you need for the rest of the week.

Questions?

Before We Continue

About Concepts

• ML vs. statistics
• When prediction matters
• Workshop structure

About Practice

• Technical requirements
• What we’ll build
• How to prepare

No question is too basic — we’re all learning together.