Spectra intrinsically possess domain knowledge, making possible a domain-based feature selection model. The random subspace method, in combination with domain-knowledge-based feature sets, leads to improved classification accuracies in real-life biomedical problems. Using such feature sets allows for an efficient reduction of dimensionality, while preserving interpretability of classification outcomes, important for the field expert. We demonstrate the utility of domain knowledge-based features for the random subspace method for the classification of three real-life high-dimensional biomedical magnetic resonance (MR) spectra.