Medication-change relation candidates¶

generate_medication_change_candidates() turns existing medication spans and explicit start, stop, hold, resume, increase, or decrease cues into reviewable relation candidates. It reuses OpenMed's deterministic local event-frame extractor and can attach a nearby event-time span.

from openmed.clinical import generate_medication_change_candidates

text = "Plan: hold aspirin today."
start = text.index("aspirin")
candidates = generate_medication_change_candidates(
    text,
    [{"label": "MEDICATION", "start": start, "end": start + 7}],
)

Each candidate carries offset-and-SHA-256 evidence for the medication, trigger, and optional time; separate medication and event assertion axes; a conflict state; and a bounded confidence. Serialized candidates do not copy source surfaces. Invalid offsets and unsupported inputs fail with value-free messages.

Every candidate has review_required=True and prescribing_action=False. Candidates never create an order, treatment plan, or recommendation. The API is deterministic, performs no network call, and requires upstream medication spans.