Independent component analysis (ICA) has the potential of automatically determining metabolite, macromolecular and lipid (MMLip) components that make up magnetic resonance (MR) spectra. However, the realiability with which this is accomplished and the optimal ICA approach for investigating in vivo MR spectra, have not yet been determined. A wavelet shrinkage de-noising based enhancement algorithm, utilising a newly derived relationship between the real and imaginary parts of the MR spectrum, is proposed. This algorithm is more robust compared with conventional de-noising methods. The two approaches for applying ICA, blind source separation (BSS) and feature extraction (FE), are thoroughly examined. A feature dimension selection method, which has not been adequately addressed, is proposed to set a theoretical guideline for ICA dimension reduction. Since the advantages and limitations of BSS-ICA and FE-ICA are different, combining them may compensate their disadvantages and lead to better results. A novel ICA approach involving a hybrid of the two techniques for automated decomposition of MRS dataset is proposed. It has been demonstrated that hybrid ICA provides more realistic individual metabolite and MMLip components than BSS-ICA or FE-ICA. It can aid metabolite identification and assignment, and has the potential for extracting biologically useful features and discovering biomarkers.