Skip to content

CRC Screening Deployment Notes

This file documents deployment considerations specific to the AIM-AHEAD CRC Screening Study.

Data Source

Patient cases were drawn from de-identified electronic health record (EHR) data, converted to OMOP CDM format. The conversion involved:

  1. EHR extraction — De-identified patient records with family history, medical history, social history, chief complaints, physical examination findings, and lab results
  2. OMOP mapping — Clinical concepts mapped to standard OMOP vocabulary codes via Athena
  3. AI score generation — A machine learning model generated CRC risk scores for each patient, stored as OMOP observations (concept ID 45614722)
  4. Quality review — Clinical team reviewed mapped data for accuracy and completeness

IRB Considerations

The CRC screening study operated under IRB approval for:

  • Use of de-identified EHR data (HIPAA Safe Harbor)
  • Clinician participants (not patients) as research subjects
  • Informed consent for clinician participation
  • Data export restricted to anonymized identifiers (user_id is SHA-256 hash of email)

When adapting AugMed for your study, consult your institution's IRB regarding:

  • Data classification — De-identified vs. limited dataset vs. identifiable
  • Participant consent — Whether clinician reviewers are considered human subjects
  • Data storage — Where AugMed is hosted and how data is protected
  • Export controls — Who can access exported research data

Clinical Context

Screening Guidelines

The CRC study was designed around USPSTF (U.S. Preventive Services Task Force) colorectal cancer screening guidelines, which recommend:

  • Regular screening for adults aged 45-75
  • Stool-based tests (FIT, FIT-DNA) as initial screening options
  • Colonoscopy as both a screening and diagnostic procedure
  • Consideration of family history and personal risk factors

Questionnaire Design Rationale

The two primary outcome questions (risk assessment and screening recommendation) were designed to:

  1. Risk assessment (1-5 scale): Capture the clinician's overall CRC risk judgment — ordinal scale enables both ordinal and continuous analysis
  2. Screening recommendation (categorical): Map directly to clinical action items in screening guidelines — enables analysis of appropriateness against guideline-concordant care

Feature Selection Rationale

The clinical features shown to participants (family history, medical history, social history, physical exam) were selected because they are:

  • Available in standard OMOP-formatted EHR data
  • Clinically relevant to CRC risk assessment per screening guidelines
  • Controllable at the individual feature level (enabling factorial experimental designs)

Production Infrastructure

The CRC study was deployed on AWS:

  • API + Frontend: AWS ECS (Fargate) behind an Application Load Balancer
  • Database: AWS RDS (PostgreSQL)
  • Domain: augmed1.dhep.org
  • Infrastructure-as-code: Terraform (see augmed-infra repository)

For simpler deployments, AugMed also supports Railway (one-click) and Docker Compose (self-hosted). See Deployment for all options.

Recruitment

Clinician participants were recruited through professional networks and clinical partnerships. The recruitment survey (Qualtrics) collected:

  • Professional role and specialty
  • Years of clinical practice
  • CRC screening experience
  • Practice location (state)

This survey data was joined with AugMed export data using participant email as the key. See Exporting Data for details on the recruitment survey integration.