Artificial Intelligence in Public Health

Machine Learning

Systems that learn and improve from experience without being explicitly programmed. ML helps departments move from descriptive reporting to predictive and prescriptive analytics.

Examples: outbreak pattern detection, immunization clinic demand forecasting, follow-up appointment prediction

Natural Language Processing

Enables computers to understand, interpret, and generate human language. NLP converts unstructured text into structured data that can be analyzed and acted upon.

Examples: automated reportable condition identification, social media monitoring, multilingual chatbots, report summarization

Predictive Analytics

Uses statistical models and machine learning to estimate the likelihood of future events based on historical and real-time data — often the most accessible form of AI for health departments.

Examples: influenza forecasting, hospital occupancy prediction, heat illness risk identification

Computer Vision

Systems that interpret and analyze visual information such as photos, scans, or video. Most applications require collaboration with environmental health or academic partners.

Examples: mosquito breeding site detection from satellite imagery, housing hazard identification, telemedicine assessment support

Generative AI

AI that learns patterns from existing data to create new content — text, images, audio, or code. Generative AI produces original outputs that can be reviewed and refined by staff.

Examples: plain-language health education materials, tailored outreach messages, draft protocols and reports, synthetic datasets

Agentic AI

Systems that can analyze data, generate content, and take a sequence of actions toward a defined goal across multiple systems, following explicit rules with human oversight.

Examples: multi-step workflow coordination, information stream monitoring and classification, routine operational task automation

Equity by Design

Monitor for disparate impact on all populations. AI projects should reduce, not widen, health disparities.

Human Oversight

Humans retain final decision authority always. AI supports decision-making — it does not replace professional judgment.

Transparency

Explainable systems with auditable outputs. Communities and partners should understand how AI is being used.

Privacy and Security

HIPAA-compliant data governance and access controls. Protecting individual privacy is non-negotiable.

Continuous Monitoring

Quarterly equity audits and performance reviews ensure AI systems continue to perform as intended over time.

Community Engagement

Meaningful engagement with affected communities in the design, testing, and oversight of AI systems.

8-Play Implementation Model

Plan

Set Direction

1 Vision and Guardrails
2 Readiness Assessment
3 Use Case Prioritization
4 Implementation Roadmap

Build

Execute and Engage

5 Stakeholder Engagement
6 Funding Alignment
7 Change Management

Govern

Monitor and Sustain

8 Monitoring and Evaluation

Phase 1: Pilot

Start small and focused to build confidence and evidence.

1 program, 1 data source, 1 target user group
90–120 day window
Establish baselines before launch
Document everything with structured debrief

Phase 2: Scale

Expand successful pilots with enhanced safeguards.

Equity impact assessment before scaling
Transition to production infrastructure
Retrain all staff who interact with AI
Update governance agreements and contracts

Phase 3: Sustain

Build for long-term operational success.

Assign named operational owner
Model review calendar and retraining schedule
Define sunset criteria for system retirement
Build sustainment costs into base budget

1

AI Readiness Self-Assessment

5-domain scoring tool for leadership, data, workforce, technology, and partnerships.

2

AI Project Charter

Defines scope, metrics, equity considerations, and risk register.

3

Equity Impact Assessment

Checklist for evaluating who is affected, bias detection, and remediation planning.

4

Implementation Timeline

Quarterly roadmap covering Discover, Pilot, Scale, and Sustain phases.

5

Change Management Plan

Communication, training, and adoption support framework.

6

Performance Dashboard

Model accuracy, equity metrics, user adoption tracking.

7

Incident Response Checklist

24-hour escalation protocol with root cause analysis categories.

8

Funding Strategy Checklists

Federal grant alignment, budget planning, and sustainability planning.

9

Data Quality Assessment

Evaluation of completeness, timeliness, accuracy, consistency, and bias.

10

Vendor Evaluation Checklist

Framework for evaluating AI vendors against public health requirements.

11

Community Engagement Tool

Planning framework for meaningful community participation in AI oversight.

Visual Overview

Playbook Framework at a Glance

AI Implementation Playbook visual framework showing the 5-Play Planning Framework, 5 Readiness Domains, 3-Phase Implementation, Federal Funding Strategy, AI Governance Framework, Key Playbook Tools, and 12-18 Month Quick-Start Roadmap

CDC Data Modernization Initiative

Foundation for AI-ready data infrastructure and interoperability. Treat AI as a capability layer on top of DMI investments.

Public Health Infrastructure Grant

Broad, flexible funding for workforce, data systems, governance, and health equity — all critical prerequisites for responsible AI.

Braided Funding Approach

Combine HRSA, SAMHSA, Title X, and state/local sources using a 4-play funding strategy: Map, Braid, Incorporate, and Sustain.

Key Findings at a Glance

72–80% reduction in surveillance chart review time using GPT-4o
6-day median lead-time advantage in outbreak detection with AI-guided inspection
AI-generated vaccination messages rated significantly more effective than CDC-authored messages (p < 0.001)
288-million-node synthetic population model quantified vaccine hesitancy impact across 50 states
89.2% source citation accuracy for AI-assisted regulatory document search

This white paper synthesizes the current evidence base for generative and agentic AI applications in U.S. governmental public health. Drawing on 21 peer-reviewed studies and technical reports published between 2020 and early 2026, it provides state, territorial, local, and tribal (STLT) public health agencies with an evidence-grounded assessment of what these technologies can do, what they cannot, and how to implement them responsibly.

The evidence demonstrates measurable benefits across six application domains: public health messaging, disease surveillance, epidemic modeling, health equity analysis, community-engaged and tribal health AI, and regulatory and administrative functions.

Public Health Messaging and Communication

GPT-3-generated COVID-19 vaccination messages were rated as having significantly stronger arguments than CDC-authored messages (p < 0.001). Across 42 U.S. counties, an AI system produced messages six times more likely to be posted by county-level agencies. A critical "source label" effect was found: use AI to draft messages, review through standard workflows, and attribute to the agency.

Disease Surveillance and Outbreak Detection

The VA's GPT-4o deployment achieved 100% specificity and 82% sensitivity while reducing per-chart review time by 72–80%. For foodborne pathogen risk, AI-guided approaches captured 1.8x more critical violations with a 6-day median lead-time advantage. The CDC's SGMC pipeline curated metadata for 2.3 million genomic accessions at 94.8% accuracy.

Epidemic Modeling and Pandemic Preparedness

The EpiHiper national-scale model (288-million-node synthetic population, 50 states, 3,300 counties) quantified vaccine hesitancy impact: faster vaccination averted 6.7 million infections and 39,400 deaths. Generative AI models consistently delivered the most accurate short-term predictions across four countries.

Health Equity and Environmental Health

MortalityMinder (open-source, GitHub) integrates 70+ social/economic factors with GPT-4, producing 636 unique infographics. A GeoAI pipeline using NASA foundation models produced hourly 10–30m resolution heat maps, capturing nocturnal heat retention in underserved neighborhoods.

Community-Engaged and Tribal Health AI

The Alaska Tribal Health System developed a community-engaged AI/ML framework using community-based participatory research for medevac utilization optimization. AI deployment must begin with community governance, not technology selection.

Regulatory and Administrative Applications

GPT-4 Turbo with RAG achieved 89.2% source citation accuracy for FDA guidance documents. State-level applications include California's AI for outbreak management, Texas's ML for Medicaid fraud detection, and Massachusetts's opioid overdose prevention system.

LLM-assisted health communication drafting (p < 0.001 effectiveness)
Open-source health equity analysis (MortalityMinder)
AI-assisted regulatory document search (89.2% citation accuracy)
LLM-assisted code generation for epidemiological modeling

Surveillance chart review augmentation (72–80% time reduction)
Foodborne risk prediction and inspection scheduling (1.8x detection lift)
Synthetic population generation for preparedness (r = 0.75)
GeoAI environmental health and heat risk mapping

National agent-based epidemic models (requires HPC)
Genomic metadata curation at scale (2.3M+ accessions)
Full agentic surveillance workflows
Community-engaged AI in tribal or underserved settings

Performance Domain	Evidence-Based Benchmark
Surveillance chart review efficiency	Target: 70%+ reduction. Benchmark: 72–80% achieved (VA GPT-4o)
Infection classification specificity	Target: ≥95%. Benchmark: 100% achieved (VA GPT-4o)
Message effectiveness	Target: At least as effective as manual. Benchmark: Significantly more effective (p < 0.001)
Document search accuracy	Target: ≥85% correct citation. Benchmark: 89.2% (GPT-4 Turbo + RAG)
Outbreak detection lead time	Target: ≥5-day advantage. Benchmark: 6-day median (foodborne AI targeting)
Equity: model disparity	No significant accuracy difference across demographic subgroups before deployment
Staff time reallocation	Document proportion of time redirected to higher-complexity work within 6 months

Human Oversight Is Non-Negotiable

AI systems require human adjudication for consequential decisions. This is a design requirement rooted in public health accountability. GPT-4o misclassified infection timing in 7 of 50 cases; GPT-4 Turbo produced incorrect regulatory information in 13.4% of responses.

Data Quality Precedes AI

Every AI application is constrained by data quality. The highest-leverage pre-AI investment is improving data quality, standardization, and integration in foundational data systems.

Equity Cannot Be Retrofitted

Models trained on non-representative data systematically underperform for underrepresented communities. Representative data collection, community engagement, and equity impact assessment must come first.

Open-Source Lowers Barriers

Resource-constrained agencies can leverage MortalityMinder, SGMC pipeline, Von Hoene synthetic population code, EpiHiper, and public data sources (BRFSS, NHIS, PLACES, NCBI SRA) without proprietary infrastructure.

Funding AI Use Cases Through Existing Grants

Funding Mechanism	Evidence-Aligned AI Applications
CDC PHEP	Epidemic modeling, synthetic population generation for surge planning
CDC OSTLTS	Surveillance modernization, LLM-assisted chart review, genomic metadata curation
HRSA	Health equity analysis (MortalityMinder, GeoAI heat risk mapping)
SAMHSA	Agentic AI for opioid overdose early identification and intervention
ACF / Tribal Health	Community-engaged AI governance (ATHS CBPR framework)
USDA / Food Safety	Foodborne pathogen risk prediction and AI-guided inspection targeting

AI for Public Health Practice

AI Technologies for Public Health