Benchmarking domain adaptation methods for cross-site antimicrobial resistance prediction from MALDI-TOF mass...

Matrix-assisted laser desorption/ionization time-of-flight mass spectrometry (MALDI-TOF MS) enables rapid species identification in clinical microbiology and shows promise for predicting antimicrobial resistance (AMR) from mass spectra. However, models trained at one hospital site suffer substantial performance degradation at another site due to differences in instruments, sample preparation, and patient populations. Despite numerous domain adaptation (DA) methods in the machine learning literature, none has been systematically benchmarked for cross-site MALDI-TOF AMR prediction. This study presents the first comprehensive benchmark evaluating 13 methods (spanning baselines, supervised transfer learning, and unsupervised DA) across five transfer scenarios involving three public datasets (DRIAMS, MS-UMG, and MARISMa) covering up to 20 species-antibiotic pairs. In total, the benchmark comprises over 15,000 experiments with five random seeds per configuration. A label-efficiency analysis across all five scenarios further examines how model performance scales with 10%, 25%, 50%, and 75% of available target-site labels. The results demonstrate that simple fine-tuning with target-site labels closes 92%-97% of the domain gap and dominates all unsupervised DA methods, which yield only 0%-6% improvement over source-only baselines. The label-efficiency analysis reveals that for competitive transfer methods, as few as 25% of target labels suffice to recover 81%-94% of full supervised performance on cross-site scenarios. These findings provide practical guidelines for clinical deployment: collecting a modest number of labeled samples at the target site is far more effective than applying sophisticated unsupervised adaptation techniques.