Artificial Intelligence Drug-Drug Interaction Screening Under Variable Herbal Co-Formulation Records: A Graph-Neural Retrieval Evaluation of Adverse-Event Signal Loss in Primary Care

Herbal co-formulation records in primary care are heterogeneous: the same botanical may appear as a single-ingredient decoction, a multi-herb formula, a proprietary blend, or a coded entry with no constituent list. Drug-drug interaction screening systems that retrieve candidate interactions from such records must decide, for each patient, whether an undocumented constituent is present. the present analysis formalises that decision as a retrieval problem over a bipartite graph of records and constituents, and derives the conditions under which a graph-neural retriever loses adverse-event signal. The central result is a bound on signal loss expressed in terms of record ambiguity, constituent co-occurrence density, and the retrieval threshold. The bound shows that signal loss is not monotone in model capacity: increasing the retriever's expressiveness reduces false negatives but amplifies false positives when co-occurrence density is high, so that the net signal can decline. A design-based evaluation is specified to characterise the operating characteristics of the retrieval estimator under a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64), with power estimated across 2000 design cells. The analysis indicates that 56 clusters are required to detect a 5-percentage-point difference at 80% power. The contribution is a formal account of why herbal-record ambiguity degrades interaction screening in a way that additional model capacity cannot remedy, and a specification of the record-standardisation conditions under which screening becomes reliable.