Influence of Basis Set Composition on Metabolite Quantification of ¹ H‐MRS at 3 T: Combining In Silico, In Vivo and In Vitro Evidence

Magnetic resonance spectra are quantified by model fitting in the time or frequency domain to a basis set consisting of metabolites present in the measured sample. Despite some work on basis set composition, it remains unclear which metabolite components should be included in the basis set models and how these impact the quantification of key metabolites such as glutamate (Glu). This lack of consensus contributes to reproducibility issues across research groups. Here, we use synthetic, human brain and phantom data to assess how basis set choice impacts quantification, focusing on the bias-variance trade-off under different SNR conditions. Simulated 1D spectra mimicking in vivo human brain data at 3 T were used to identify basis models that minimise bias and variance for our four key metabolites of interest: Glu, creatine (tCr), choline-containing compounds (tCho) and N-acetylaspartate (tNAA) under increasing noise levels. This informed analyses of human brain and phantom data collected at 3 T over 81 and 120 min, respectively, using PRESS. We find that basis set composition significantly affected Glu, tCr, tCho and tNAA concentrations. Specifically, the inclusion of γ-aminobutyric acid, glutathione, N-acetylaspartylglutamate and glucose improved Glu quantification, achieving bias and variance below 10%. Including partner metabolites for tCr and tNAA (phosphocreatine and N-acetylaspartylglutamate) offered no significant benefit. In contrast, tCho quantification remained inconsistent likely due to spectral overlap. We show that minimal basis set models provide accurate quantification while reducing variance. However, the number of metabolites accurately modelled depends on data quality and SNR. High-SNR spectra enable the inclusion of additional metabolites, while low-SNR data risk overfitting. Differences in metabolite concentrations reported in the literature may partly reflect variations in prior knowledge models, emphasising the need for clear descriptions of analysis methods in MRS research.