Abstract
BACKGROUND/OBJECTIVES: The progression of antibiotic resistance is increasingly recognized as a dynamic and time-dependent phenomenon, challenging conventional diagnostics that define resistance as a binary trait. METHODS: Biomolecules have fingerprints in Fourier-transform infrared spectroscopy (FTIR). The targeting of specific molecular groups, combined with principal component analysis (PCA) and machine learning algorithms (ML), enables the identification of bacteria resistant to antibiotics. RESULTS: In this work, we investigate how effective classification depends on the use of different numbers of principal components, spectral regions, and defined resistance thresholds. Additionally, we explore how the time-dependent behavior of certain spectral regions (different biomolecules) may demonstrate behaviors that, independently, do not capture a complete picture of resistance development. FTIR spectra were obtained from Staphylococcus aureus exposed to azithromycin, trimethoprim/sulfamethoxazole, and oxacillin at sequential time points during resistance induction. Combining spectral windows substantially improved model performance, with accuracy reaching up to 96%, depending on the antibiotic and number of components. Early resistance patterns were detected as soon as 24 h post-exposure, and the inclusion of all three biochemical windows outperformed single-window models. Each spectral region contributed distinctively, reflecting biochemical remodeling associated with specific resistance mechanisms. CONCLUSIONS: These results indicate that antibiotic resistance should be viewed as a temporally adaptive trajectory rather than a static state. FTIR-based biochemical profiling, when integrated with ML, enables projection of phenotypic transitions and supports real-time therapeutic decision-making. This strategy represents a shift toward adaptive antimicrobial management, with the potential to personalize interventions based on dynamic resistance monitoring through spectral biomarkers.