Atomic summary claim segmentation¶

OpenMed can split a post-de-identification summary into deterministic claim spans before citation or natural-language-inference checks. Every emitted claim retains an exact half-open character offset into the supplied summary.

from openmed.clinical.summary_claim_segments import segment_summary_claims

summary = "Signal alpha improved, but marker beta persisted."
result = segment_summary_claims(summary)

for claim in result.segments:
    assert summary[claim.start : claim.end] == claim.text
    print(claim.offset, claim.review_required)

The implementation is local and deterministic. It uses the package's pinned sentence segmenter, makes no network request, loads no model, and adds no telemetry. Semicolons, newlines, and sentence endings are stable boundaries. For English summaries, commas, and, but, and yet split only when both sides have an explicit subject and finite predicate.

Review gate¶

Conservative segmentation is deliberate. A clause that still contains a coordination, alternative, causal, temporal, or relative-clause marker is kept intact and receives:

review_required = true
review_reason = "unsegmentable_compound"

Callers must stop before claim-level citation and NLI checks until a reviewer resolves that span. Non-English summaries retain deterministic sentence-level offsets but receive unsupported_language, because the English clause rules cannot establish that those spans are atomic.

Privacy boundary¶

Pass only summary text produced after de-identification. SummaryClaimSegment contains its exact summary slice because downstream verification needs it, but repr(), to_dict(), and to_json() omit claim text and expose only offsets, counts, language, schema version, and controlled review metadata. Exceptions use fixed messages and do not echo submitted values. This is a verification aid, not a compliance certification or autonomous clinical decision system.