Diagnosis-to-treatment relation candidates¶

generate_diagnosis_treatment_candidates() requires explicit linking language between existing diagnosis and treatment spans. Supported guarded patterns include diagnosis-first forms such as “treated with” and treatment-first forms such as “for” or “to treat.” Co-occurrence alone produces no candidate.

from openmed.clinical import generate_diagnosis_treatment_candidates

text = "Plan: possible pneumonia treated with ceftriaxone."
candidates = generate_diagnosis_treatment_candidates(
    text,
    [
        {"label": "DIAGNOSIS", "start": 15, "end": 24},
        {"label": "MEDICATION", "start": 38, "end": 49},
    ],
    max_distance=80,
    allowed_sections={"plan"},
)

Both endpoints must share a deterministic sentence and section, satisfy max_distance, and—when supplied—belong to allowed_sections. Results include offset/hash cue evidence, assertion axes for both endpoints, and a controlled none, possible, conditional, or refuted uncertainty state.

Every output has review_required=True, treatment_recommendation=False, and treatment_evaluated=False. The API does not recommend or assess treatment, copies no source surfaces into serialization, is deterministic, and performs no network call.