Auris: End-to-End Pipeline for Structured FHIR Resource Extraction from Ambient Voice
Auris — End-to-End Ambient Voice-to-FHIR Pipeline with Structured Resource Validation
Summary
This paper introduces 'Auris,' an end-to-end pipeline that generates validated FHIR R4 resources directly from clinician-patient audio. The system moves beyond simple transcription to achieve structured data extraction, providing information in a format usable by Electronic Health Records (EHRs).
Details
This research manuscript proposes the 'Auris' pipeline. Its goal is to generate structured and validated FHIR R4 resources (such as Observation, Condition, MedicationRequest) directly from clinician-patient audio dialogues, rather than merely producing narrative SOAP notes. Technically, it utilizes Whisper-large-v3 for transcription, pyannote 3.x for speaker diarization, and JSON-Schema constrained decoding for information extraction. This process ensures that the resulting data conforms to FHIR profiles. Historically, ambient AI scribing often produced unstructured narrative SOAP notes, requiring downstream parsing by EHRs. Auris addresses this limitation by making structured FHIR resources the primary output. Furthermore, drawing on insights from Adedeji et al., the pipeline is designed with robustness against ASR failure modes (like drug-name mistranscription), ensuring that ambiguous transcriptions are not silently integrated into the final FHIR resources.
Original content copyright by respective publishers