Work

Silence Is Safety: A Calibrated Clinical NLI Verifier for MedNLI

Year: 2026

Type: preprint

Abstract: This preprint presents AetherMind MedNLI Clinical NLI Verifier, a calibrated clinical natural-language inference model for the MedNLI benchmark. The model performs three-way inference over clinical premise-hypothesis pairs: entailment, neutral, and contradiction. On the MedNLI test split, the verifier achieves 88.19% accuracy with 95% confidence interval [86.50 89.87], macro F1 of 88.21%, contradiction recall of 92.41%, and contradiction precision of 93.99%. Temperature scaling fitted on the development split improves expected calibration error from 0.090 to 0.037. The paper also evaluates selective abstention for safer clinical claim verification workflows.

Source: Zenodo (CERN European Organization for Nuclear Research)

Author: Najm, Sameer

Language: English

Topic: Machine Learning in Healthcare

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

SDG: Good health and well-being

Open Access status: green

OpenAlex ID: W7164672124

Cites:

Cited by:

Related to: