Assertion-aware clinical NLI pairs¶

Clinical natural-language inference must keep the source assertion state next to the premise and hypothesis. A negated, uncertain, or conditional finding is not equivalent to affirmed evidence, even when the surface concept is the same. OpenMed therefore requires complete assertion metadata on both sides of every constructed pair.

These pairs are deterministic assistive inputs for qualified human review. They do not diagnose, recommend treatment, or make an autonomous clinical decision.

Construct a pair¶

Use the existing ClinicalAssertion axes or a mapping with equivalent fields. The required axes are:

Both the premise and the hypothesis must provide these axes. Missing metadata is rejected rather than defaulted. If redundant forms are supplied, they must agree; for example, temporality="recent" with hypothetical=True is rejected.

from openmed.clinical import build_nli_pair

pair = build_nli_pair(
    {
        "text": "synthetic source finding",
        "assertion": {
            "negation": "negated",
            "uncertainty": True,
            "hypothetical": False,
        },
        "start": 12,
        "end": 35,
    },
    {
        "text": "synthetic target claim",
        "assertion": {
            "negation": "affirmed",
            "certainty": "certain",
            "temporality": "recent",
        },
    },
)

pair.premise_assertion and pair.hypothesis_assertion expose the normalized axes. The pair does not infer an NLI label: a downstream local model or rule-based verifier must retain these fields when interpreting its result. Every pair is marked requires_clinician_review=True.

Privacy boundary and offline behavior¶

The pair retains premise and hypothesis text only for the caller-controlled model-input boundary:

model_input = pair.to_model_input()
# model_input keeps premise_assertion and hypothesis_assertion beside the text.

For a backend that accepts only a text tuple, use pair.to_text_pair() and retain the assertion sidecars from the pair next to the backend result.

to_dict(), to_audit_dict(), to_json(), and repr(pair) never include raw text. They contain only text hashes, lengths, optional source offsets, the assertion axes, a deterministic pair identifier, and the review advisory. Validation exceptions identify only the side and field that failed; they do not echo the supplied text or invalid value. Keep raw text inside the local review boundary and use the audit representation for logs, reports, and persisted artifacts.

Construction performs no model loading, filesystem access, network call, logging, telemetry, or remote provider lookup. Batch construction preserves input order and produces the same pair identifiers and JSON for the same inputs.

Limitations¶

Assertion metadata is an evidence safeguard, not a clinical truth guarantee. The ConText layer and any upstream extractor remain responsible for producing the metadata. A qualified reviewer must inspect source evidence and the model or rule-based NLI result before downstream use. This feature does not bundle clinical corpora, restricted terminology, credentials, or proprietary model services.

Validation¶

The focused suite uses synthetic values only and runs offline:

.venv/bin/python -m pytest tests/unit/clinical/test_nli_assertion_pairs.py -q

When the optional checkout-local virtual environment is unavailable, the equivalent command is python3 -m pytest ....