CHIA Grant Architecture Walkthrough

Seattle Children's Hospital — Pediatric Epilepsy Research Platform

re:Invent 2025 Enhanced
UC1: Data Lake
UC2: AI Neuroimaging
Combined
Security & Governance
Cost Summary
UC1: Pediatric Epilepsy Data Lake Architecture
Unified health data platform integrating EHR, imaging, genomics, and wearable data for epilepsy research
Existing AWS Service
re:Invent 2025 — NEW
Data Sources
💗
Epic FHIR
EHR Records
📡
PACS / DICOM
MRI, EEG Images
⏱️
Wearables
IoT Seizure Data
🧬
Genomics
VCF/CRAM Files
Ingestion
❤️
HealthLake
FHIR Data Store
🧠
HealthImaging
Medical Images
🧬
HealthOmics
Genomic Workflows
NEW
⚡
Lambda Durable
Long-running ETL
Storage
🗄️
Amazon S3
Data Lake (HIPAA)
NEW
📍
S3 Vectors
Embeddings Store
🛡️
Lake Formation
Governance
Processing
🔀
AWS Glue
ETL & Catalog
🧠
SageMaker
HyperPod Cluster
NEW
💾
Checkpointless
Training Resilience
Analytics & AI
🔍
Athena
SQL Analytics
✨
Bedrock
Foundation Models
NEW
✨
Nova Forge
Custom Model Fine-tune
NEW
🤖
AgentCore
Autonomous Agents
Consumption
📊
QuickSight
Dashboards
🌐
API Gateway
Research APIs
▦
Research Portal
Clinician UI
UC2: AI-Enhanced Neuroimaging Pipeline
Super-resolution MRI enhancement model trained on multi-institutional data, deployed as custom Bedrock model
Existing Service
re:Invent 2025 — NEW
Research Environment
📁
Public Datasets
HCP, fastMRI, BraTS
🏢
UPMC Collaboration
Paired MRI Dataset
⚙️
Preprocessing
DICOM → NIfTI, QC
🗄️
S3 Research Bucket
Curated Training Data
Training Environment
💻
HyperPod Cluster
p4d.24xlarge × 4
NEW
💾
Checkpointless Training
-10% training cost
NEW
💻
Trainium3
Next-gen ML silicon
📈
W&B / MLflow
Experiment Tracking
Production (Inference)
✅
Model Evaluation
PSNR, SSIM, Clinical
✨
Custom Model Import
Bedrock Inference
🖥️
Inference Endpoint
Real-time Enhancement
🩺
Clinical Integration
PACS Viewer Plugin
🔀 Data Flow: UC1 ↔ UC2 Integration
UC1: Data Lake
Neuroimaging + EHR
→
S3 Staging
MRI DICOM Export
→
UC2: Model Training
Super-Resolution
→
Enhanced Images
4× Resolution
→
UC1: Data Lake
Research Corpus

The Data Lake (UC1) provides neuroimaging data to train the AI model (UC2). Enhanced images flow back into the lake, enriching the research corpus and enabling clinician access via the Research Portal.

🔗 Shared Infrastructure Services
🗄️
Amazon S3
Data Lake
🔑
AWS KMS
Encryption
📜
CloudTrail
Audit
🛡️
Lake Formation
Governance
📊
Security Hub
Compliance
🌐
VPC + PrivateLink
Networking
👤
IAM + SSO
Identity
📊
CloudWatch
Monitoring
⚡ re:Invent 2025 Enhancements (Shared)
NEW
📍
S3 Vectors
Replaces OpenSearch — $500/mo savings, simpler architecture
NEW
💾
Checkpointless Training
10% compute savings — resilience without checkpoint overhead
NEW
💻
Trainium3
40% inference cost reduction — purpose-built for ML
NEW
⚡
Lambda Durable + AgentCore + Nova Forge
ETL orchestration, autonomous agents, custom fine-tuning
Security & Governance — HIPAA-First Architecture
Defense-in-depth controls enabled from Day 1 of the POC. All services are HIPAA-eligible with BAA coverage.
🔐 Encryption & Access Control
🔑
AWS KMS
Customer-managed CMKs, automatic key rotation every 365 days
🛡️
Lake Formation
Tag-based access: per-site, per-study, column-level PHI masking
👤
IAM + Identity Center
Least-privilege, federated SSO with SCH AD, no long-lived creds
🌐
VPC + PrivateLink
All PHI traffic on private network, no internet-facing endpoints
🔍 Detection & Monitoring
🔎
Amazon Macie
Continuous S3 scanning for PHI leakage, alerts on unencrypted PII
⚠️
Amazon GuardDuty
Threat detection: anomalous API calls, credential compromise, exfiltration
📜
AWS CloudTrail
Full API audit trail, 7-year retention for compliance
📊
AWS Security Hub
HIPAA conformance pack, continuous compliance scoring
🤖 AI Governance
🤖
AgentCore Policy (Cedar)
Deterministic enforcement: default-deny, forbid-wins-over-permit, automated reasoning
🛡️
Bedrock Guardrails
Content filtering, PII redaction, topic denial — applied regardless of model
Policy Authoring
Cedar policies define what AI agents can access. Automated reasoning validates policies are well-formed before deployment.
Runtime Enforcement
Default-deny, forbid-wins-over-permit. Every agent action evaluated against governance policies in milliseconds. Deterministic — not probabilistic.
Audit & Compliance
Full decision log for every agent action — approved or denied. Maps directly to AI Review Board governance requirements.
💡 For the AI Review Board: Cedar policies are the technical enforcement layer for the governance rules your board defines. "No AI agent may access genomic data without IRB approval tag" becomes a Cedar policy that is enforced deterministically at the gateway.
📅 Security Deployment Timeline
Week 1-2
Security baseline: KMS, VPC, IAM roles, CloudTrail, Security Hub HIPAA pack, GuardDuty
Week 3-4
Data governance: Lake Formation tags, Macie scanning, de-identification pipeline validation
Week 5-8
AI governance: AgentCore policies, Bedrock Guardrails, model access controls
Week 9-12
Penetration testing, compliance audit, AI Review Board sign-off, production readiness
UC1 — Year 1 Estimate
$58,000
Covered by CHIA credits
UC2 — Year 1 Estimate
$66,000
Covered by CHIA credits
Combined Total
$124,000
~$26K headroom on $150K award
CHIA Award Coverage
$150,000
AWS credits — full coverage ✓

Year 1 Cost Breakdown

UC1 + UC2 Combined = ~$124K (CHIA Award: $150K)

UC1: Data Lake
$58K
UC2: Neuroimaging
$66K
Combined
$124K
CHIA Award $150K
Storage Compute Platform AI/ML

re:Invent 2025 Cost Savings

Waterfall: Original → Optimized

$140K
Pre-Optimization
-$6K
S3 Vectors
-$4.2K
Checkpointless
-$5.8K
Trainium3
$124K
Post-Optimization

re:Invent 2025 Cost Savings

S3 Vectors
Replaces OpenSearch Serverless
-$500/mo
Checkpointless Training
Eliminates checkpoint I/O
-10% compute
Trainium3 Inference
vs GPU-based inference
-40% inference

Post Year 1: Steady-State Operations

Monthly Operational
$8-10K
After POC optimization
Annual Steady-State
~$100K
With Savings Plans
Comparison
<1 FTE
Less than one research coordinator salary
Year 2+ cost drivers: Storage grows ~20%/year as cohort expands. Compute stays flat (inference only, no retraining). Savings Plans (1-year commit) reduce compute 30-40%. Right-sizing after POC usage patterns are established typically saves an additional 15-20%.
💡 Budget context: $100K/year supports a platform serving 6,000+ epilepsy patients across 30+ PERC sites — less than the cost of one FTE research coordinator doing manual chart abstraction.

Cost of Inaction — What Manual Alternatives Cost

Manual Chart Abstraction (UC1)
$450K+/yr
3 FTE research coordinators × $85K salary + benefits
6,000 patients × 4 modalities = 24,000 records to abstract
Timeline: 18-24 months vs. 12 weeks with automation
Commercial 7T Scanning (UC2)
$1.5M+
500 research scans × $3,000/scan = $1.5M
Plus: scheduling delays (2-4 weeks), travel to 7T site
Not validated for pediatric routine use
✅
AWS Platform: $124K Year 1 (covered by CHIA credits)
Automated NLP + AI imaging enhancement — 10× faster, 10× cheaper than manual alternatives

Next Steps

1
Confirm genomic data format
WGS vs WES vs targeted panel — biggest cost variable for storage planning
2
Align on de-identification approach
Safe Harbor vs Expert Determination — determines FHIR export pipeline design (needs IRB input)
3
Schedule technical kickoff — Week of May 19
90-minute session with research informatics + engineering. 12-week clock starts → POC results by mid-June for CHIA submission (July 1 deadline).
What it does
Why it matters for epilepsy research
Estimated Cost