Artificial Intelligence Optimal Transport Alignment of Multilingual Clinical Triage Notes Under Variable Translation Quality: A Calibration Evaluation of Healthcare Triage Safety

Multilingual clinical triage notes are routinely translated before algorithmic triage models consume them, yet translation quality varies across languages, note types and providers, and the resulting distributional shift can silently degrade triage safety. The framework treats each translated note as a probability measure over a clinical feature space and aligns source and target measures through entropic optimal transport with a translation-quality-dependent cost. A calibration bound is derived that links the Wasserstein discrepancy between translated and reference note distributions to the expected triage-safety loss, and the bound is decomposed into a translation-noise term and an alignment term. The analysis is theoretical and methodological: no observational data are used or claimed. The framework proposes a quality-weighted Sinkhorn alignment procedure and specifies how calibration should be evaluated across translation-quality strata. The contribution is a formal characterisation of when optimal transport alignment preserves triage calibration and when it does not, together with a set of testable propositions that link translation-quality heterogeneity to triage-safety risk. The framework implies that alignment without quality weighting can increase rather than reduce triage-safety loss when translation quality is heterogeneous across languages.