Display omitted •Multivariate data analysis of SERS maps enabled MTX prediction in patient samples.•More accurate quantification with multivariate vs. univariate data analysis.•The model can be built directly in commercial serum and applied to patient samples.•The genetic algorithm identified specific wavenumbers from complex SERS spectra.•Image threshold segmentation was applied to select relevant map pixels. Despite the technological development in Raman instrumentation that has democratized access to 2D sample scanning capabilities, most quantitative surface-enhanced Raman scattering (SERS) analyses are still performed by only acquiring a single or a few spectra per sample and performing univariate data analysis on those. This strategy can however reach its limit when analytes need to be detected and quantified in complex matrices. In that case, surface fouling and competition between the target analyte and interfering compounds can impair univariate SERS data analysis, underlining the need for a more in-depth data analysis strategy based on exploiting of full-spectrum information. In this paper, a multivariate data analysis strategy was developed, for analyzing SERS maps of methotrexate (MTX) from patient samples, including all steps from baseline correction, selection of wavelength, and the relevant pixels in the map (image threshold segmentation), as well as quantitative model construction based on partial-least squares regression. Among the different baseline correction methods evaluated, standard normal variable transformation and Savitzky-Golay smoothing proved to be more suitable, while the genetic algorithm wavelength screening method was able to screen out MTX-related SERS spectral regions more efficiently. Importantly, with the here-developed process, it was sufficient to use MTX-spiked commercial serum when building quantitative models, removing the need to work with MTX-spiked patient samples, and consequently enabling time- and resource-saving quantitative analyses. Besides, the developed multivariate data analysis approach showed superior performances compared with univariate analysis, with 30 % improved sensitivity (detection limit of 5.7 µM), 25 % higher reproducibility (average relative standard variation of 15.6 %), and 110 % better accuracy (average prediction error of −10.5 %).