Patient-intake application
Three synthetic patient journeys
Tested the local intake workflow through safety checks, nurse review and FHIR R4 integration with a local HAPI test server.
Selected project · 2026
CareFlow
AI-assisted patient intake, triage and clinical workflows
Turning a patient’s intake into a structured record, with safety checks and a nurse’s decision along the way. Alongside the application, a separate model workstream explores triage prediction when vital signs are missing.
01 / The application
From intake to a reviewed record.
The workflow connects local language-model processing, explicit safety rules, nurse review and a structured FHIR record.
Patient intake
Local Qwen and LangGraph process the patient intake into structured information.
Safety checks
Rule-based checks run alongside the model-assisted workflow.
Nurse review
A nurse can confirm or override the result before it moves forward.
FHIR write-back
The reviewed information is written to a local HAPI test server using FHIR R4.
Workflow overview · Tested with synthetic patient journeys.
My contribution
I implemented the local patient-intake workflow, connected the safety checks and nurse review, and integrated FHIR R4 write-back.
The engineering focus
Make the transition from model output to a reviewed, structured record explicit, while retaining the nurse’s ability to confirm or override.
02 / The model
Laya: triage with incomplete inputs.
A separate model workstream within CareFlow, focused on prediction when some vital signs are unavailable.
I fine-tuned the ModernBERT-based Laya model in PyTorch on Korean Triage and Acuity Scale (KTAS) data, simulating missing vital signs during training.
I evaluated predictions on held-out cases against a tabular baseline. This model evaluation is separate from the Qwen and LangGraph patient-intake workflow above.
- Model and framework
- ModernBERT-based Laya · PyTorch
- Data and input conditions
- KTAS data · Simulated missing vital signs
- Evaluation approach
- Held-out cases · Comparison with a tabular baseline
03 / Evaluation
Workflow tests and model evaluation.
Application testing follows the patient journey. Model evaluation examines predictions under defined input conditions.
Laya model workstream
191 held-out KTAS records
Evaluated the fine-tuned model against a tabular baseline while examining the effect of missing vital signs.
Evaluation used synthetic patient journeys and offline KTAS data, with FHIR integration exercised against a local HAPI server.