CIRCA: Clinical Intent Representation, Cross-corpus Annotations

Companion site for the paper "Clinical Intent Extraction: A FHIR-Aligned Representation and the CIRCA Benchmark" by Alexander Apartsin (Holon Institute of Technology) and Yehudit Aperstein (Afeka College of Engineering).

Read the paper

• HTML — circa.html (source of truth; renders in any browser)
• PDF (single-column) — circa_1col.pdf
• Word (single-column) — circa_1col.docx

Supplementary material

• dataset_examples.html — What the text and native annotation of each source corpus (CLIP, MedDec, ap_parsing, PaniniQA, SIMORD/ACI-Bench) look like.
• circa_litreview.md — Literature review notes informing the paper's positioning.
• human_studies_protocol.md — Protocol for the two minimal human studies (exhaustive whole-note validation; second-annotator reliability).
• circa_forword.html — Word-conversion source used to build the DOCX (identical content to circa.html with figures rasterized to PNG).
• scripts/docx_sync.py — Post-processor that syncs Figure 1 layout and Table 7 styling in the DOCX to match the HTML.

What CIRCA is

• The Clinical Intent Extraction (CIE) task: given a clinical note, extract every prospective clinical action (follow-up, order, referral, instruction) as a structured record.
• The Clinical Intent Representation (CIR): a FHIR-aligned schema that decomposes each action into an action verb, type, coded target, timing, condition, plus two axes prior clinical-note datasets do not jointly represent — request-intent (proposal / plan / order / option, aligned to HL7 FHIR RequestIntent) and modality (a seven-valued clinical-strength taxonomy).
• The CIRCA dataset: 10,011 harmonized intents over 1,185 notes and 942 MIMIC-III patients plus the public ACI-Bench dialogue corpus, harmonized from five heterogeneous source corpora (CLIP, MedDec, ap_parsing, PaniniQA, SIMORD / ACI-Bench).

Dataset deposit

• Zenodo: 10.5281/zenodo.22251272
• Package (annotations + KART mappings + code, no source note text): circa-cir-v1.zip

The MIMIC-derived layers ship as stand-off annotations (character offsets + SHA-256) that a credentialed user rehydrates from their own MIMIC-III copy with the included rebuild.py; only the public ACI-Bench layer is distributed with its note text. See the deposit's README.md and DATASHEET.md for the reconstruction procedure and the data-provenance / compliance statement.

Citation

Apartsin A, Aperstein Y. Clinical Intent Extraction: A FHIR-Aligned Representation
and the CIRCA Benchmark. 2026.
Dataset: https://doi.org/10.5281/zenodo.22251272

Compliance

CIRCA never redistributes MIMIC-III text. Model inference during annotation ran on AWS Bedrock under a HIPAA Business Associate Agreement with a zero-retention configuration; the source-corpus data-use agreements govern any reconstructed text.