Model Card Builder
Review your Draft
AI Products
Intended Use
Patient Population
Inputs and Outputs
Performance Evidence
Safety and Limitations
Bias / Fairness
Human Factors / Workflow
Security / Privacy
Regulatory / Deployment
Company Name
Product Name
Product Version
Product URL
Product Summary
AI-assisted triage tool that helps emergency nurses prioritize incoming patients.
AI / Model Type
Large language model
Product Status
Dev
Foundation Model
Third-party API
Release Date
Next
arrow_forward
Primary intended use
Prioritize incoming emergency department patients by acuity at triage.
Intended users
Nurses
Physicians
Deployment settings
Emergency department
Inpatient
Clinical workflow
Runs when a triage assessment is saved in the EHR and shows a suggested acuity level.
What should the user do after receiving the output?
Review the suggested acuity and confirm or change it before assigning a bed.
Out-of-scope uses
Pediatric emergency care and prehospital settings.
Next
arrow_forward
Adult, pediatric, or both?
Adult
Age range
Clinical conditions or population
Adults presenting to the emergency department with undifferentiated complaints.
Geographic restrictions
Populations not evaluated, or where performance is uncertain
Pregnant patients and patients with limited English proficiency.
Next
arrow_forward
Input information types
Demographics
Vitals
Clinical notes
Where does input data come from at runtime?
Pulled from the EHR when the triage assessment is saved.
How does the product receive data?
FHIR R4 API
SMART on FHIR app
EHRs you currently integrate with
Epic
Oracle Health (Cerner)
Output type
Risk score
Recommendation
Describe the output a user sees
An acuity level from 1 to 5 with a short rationale.
Is the output intended to influence a clinical decision?
Supports a clinician who makes the decision
Data used to develop or train the model
De-identified emergency visits from three health systems, 2019 to 2023.
Next
arrow_forward
Has the product's performance been evaluated?
Yes, using clinical data
Evaluation type
Retrospective
External
What was the evaluation trying to show?
Show the suggested acuity agrees with nurse-assigned acuity.
Primary metrics and results
Agreement: 91.2% (95% CI 89.5-92.8) Under-triage rate: 2.1%
How should these results be interpreted?
Agreement is comparable to nurse-to-nurse agreement.
Comparator or baseline
Dataset description
42,000 adult emergency visits held out from model development.
Sample size
Evaluation population
Evaluation setting
Link to supporting evidence
Next
arrow_forward
Known limitations
Less accurate when the triage note is very short or missing.
Known failure modes
Under-triages atypical chest pain presentations.
When should users not rely on the AI?
The patient is unstable or vitals are missing.
Current safeguards
Nurse confirmation is required before any acuity is assigned.
Is human review of the output required?
Always
Your assessment of clinical risk level
Moderate
Basis for that assessment
Have safety or reliability been tested?
Yes
What did the safety testing try to show?
Confirm that unsafe under-triage is rare.
Safety and reliability results
Critical under-triage: 0.4%
Test type
Testing data description
1,200 cases reviewed by two emergency physicians.
Interpretation
Rate is within the range accepted by the clinical safety board.
Link to supporting evidence
Next
arrow_forward
Has performance been evaluated across patient subgroups?
Yes
Groups evaluated
Age
Sex
Race / ethnicity
Subgroup metrics and results
Agreement by subgroup, difference from overall
What did you find?
Agreement stayed within 2 points of the overall rate in every subgroup.
Testing data description
The same 42,000 held-out visits.
Populations with uncertain or insufficient evaluation
Patients over 90 had too few cases to report.
What do you do to reduce bias?
Training data was rebalanced across age groups.
Next
arrow_forward
How is the output presented to the user?
A banner in the triage screen with the acuity level and rationale.
Can users override, reject, or modify the output?
Accept, reject or modify
What happens when the AI cannot produce an output?
The banner is hidden and triage continues as usual.
Has it been evaluated inside the intended clinical workflow?
Yes
Describe the workflow evaluation
Four-week pilot in one emergency department with 18 nurses.
How do users understand, question, or give feedback on outputs?
Each suggestion lists the inputs that most influenced it.
Next
arrow_forward
Does the product process PHI?
Yes
Does it process other PII?
Yes
Does it store patient data?
No
Encryption at rest?
Yes
Encryption in transit?
Yes
Role-based access controls?
Yes
Audit logging?
Yes
Certifications and attestations
SOC 2 Type 2
ISO 27001
HIPAA security risk assessment
Next
arrow_forward
Subject to regulatory oversight?
Yes
Regulatory pathway
FDA 510(k)
Status
Cleared / authorized
Device classification
Clearance / authorization identifier
Clearance date
Current deployment status
Pilot
Where is the product available?
Does your product already support FHIR R4?
Yes
How is the model maintained and updated?
Quarterly model updates with a regression test before release.
How is performance monitored after deployment?
Monthly drift checks and nurse-reported issues.
Funding source
Self-funded
Stakeholders consulted during development
Nurses
Physicians
Health system leaders
Third-party components or data
Hosted language model API from a third-party provider.
Peer-reviewed publications
None yet
Clinical trial registration
Reimbursement status
Patient consent or disclosure
Patients are told AI supports triage in the registration notice.
Next
arrow_forward