Unlabeled Spectral Shift Informs Target-Label Acquisition Strategy Selection for Cross-Site MALDI-TOF Antimicrobial Resistance Prediction

Background/Objectives: Machine learning applied to MALDI-TOF mass spectra can flag antimicrobial resistance days before conventional susceptibility testing, but models rarely retain accuracy in a new laboratory. We asked whether this failure can be repaired without target-site labels and, if a few local labels are obtainable, how they should be chosen. Methods: We assembled 72 cross-site species–antibiotic transfer tasks from DRIAMS (eight species, 18 agents, three site pairs) in 12 prespecified clusters. After benchmarking unsupervised adaptation, five label-free shift descriptors were used to choose between Random and Hybrid acquisition at 20, 30 and 50 target labels, under nested leave-one-cluster-out validation with cluster-level sign-flip inference. Results: Source–target separability was uniformly high (domain AUROC: 0.966–1.000), while external PR-AUC varied widely (0.049–0.920), and no unsupervised method restored transportability. The selective policy numerically exceeded Always-Random at all budgets (+0.0084, +0.0078 and +0.0048 PR-AUC; p = 0.064, 0.012 and 0.045). The 30-label result was robust to excluding development tasks, duplicated phenotypes and post hoc descriptor choice, but not to accounting for isolates shared between clusters (p = 0.06–0.09), and the policy did not outperform Always-Hybrid. Conclusions: Site identity persists in MALDI-TOF spectra despite harmonization, so cross-site deployment is better planned around a small local labeling budget than around label-free correction. Shift-informed acquisition is a methodological proof of concept: its gains were small and statistically fragile, it did not outperform Always-Hybrid, and independent laboratory validation and local recalibration are required before any clinical application.