Artificial Intelligence Triage for Snakebite Envenoming Severity Under Antivenom Batch Heterogeneity: A Calibration-Transfer Evaluation of Coagulation Assay Thresholds

Antivenom batch heterogeneity poses a persistent threat to the reliability of artificial intelligence (AI) triage systems for snakebite envenoming, particularly when decision thresholds are calibrated on coagulation assay data from one antivenom production batch and deployed against another. The framework treats coagulation assay thresholds as transferable parameters and derives bounds on classification error under batch-specific measurement offsets. Using a theoretical analysis of threshold displacement, the paper establishes conditions under which calibration transfer fails and proposes a batch-robust recalibration procedure. The analysis identifies that even modest batch-to-batch variation in assay sensitivity can produce substantial misclassification of severe envenoming cases, with the magnitude of error depending on the interaction between threshold location and the distribution of assay values in the target population. The framework provides a basis for regulatory evaluation of AI triage systems across antivenom production cycles and offers a methodological template for calibration-transfer assessment in resource-constrained clinical settings.