Artificial Intelligence Triage for Postpartum Hemorrhage Risk Under Variable Blood-Loss Estimation Practices: A Calibration-Transfer Evaluation of Transfusion-Trigger Thresholds

Postpartum hemorrhage remains a leading cause of maternal mortality, and artificial intelligence triage models that recommend transfusion are increasingly proposed to support clinical decisions in low-resource settings. These models are typically trained on visually estimated blood loss, yet visual estimation is known to be systematically biased and highly variable across clinicians and institutions. The problem is framed as a calibration-transfer problem: the mapping from true blood loss to recorded blood loss differs between practices, so a model learned under one mapping encounters a shifted input distribution at deployment. A formal framework is developed that separates the physiological signal from the estimation error, defines the calibration-transfer gap, and derives conditions under which a fixed threshold produces systematically different transfusion recommendations across practices. The analysis is theoretical and methodological; no observational data are used or claimed. The framework yields three results: a bias-variance decomposition of the transfer gap, a threshold-shift condition that quantifies how much a known estimation bias moves the effective trigger, and a sensitivity bound showing that the safety margin required for transfer grows with the variance of the estimation practice. The contribution is a precise characterisation of when calibration transfer fails and a set of design conditions under which an artificial intelligence triage model can be deployed across heterogeneous blood-loss estimation practices without silent erosion of its transfusion-trigger safety margin.