Abstract
Non-targeted methods (NTMs) using matrix-assisted laser desorption/ionization time-of-flight mass spectrometry (MALDI-TOF MS) show promise in bacterial resistance detection, yet temporal variations in spectral features pose significant challenges. These proteomic patterns, which characterize bacterial phenotypes and pathological functions, may vary over time due to bacterial adaptation, virulence, or resistance mechanisms, resulting in large prediction uncertainties and potentially degrading NTM performance. We present a comprehensive screening method to detect temporal changes in MALDI-TOF spectral patterns, demonstrated using methicillin-resistant and -susceptible Staphylococcus aureus (MRSA/MSSA) isolates collected over several years. Our approach combines convolutional neural networks (CNNs) with statistical methods, including significance testing, kernel density estimation, and receiver operating characteristics for dataset shift detection. We employ Gradient-weighted Class Activation Mapping (Grad-CAM) for post hoc feature description, enabling biochemical characterization of temporal changes. This analysis reveals crucial insights into the dynamic relationship between spectral data patterns over time, addressing key challenges in developing robust NTMs for routine applications.