This work presents a first attempt to apply fuzzy cluster analysis (FCA) to analyzing stellar spectra. FCA is adopted to categorize line indices measured from LAMOST low-resolution spectra, and automatically remove the least metallicity-sensitive indices. The FCA-processed indices are then transferred to the artificial neural network (ANN) to derive metallicities for 147 very metal-poor (VMP) stars that have been analyzed by high-resolution spectroscopy. The FCA-ANN method could derive robust metallicities for VMP stars, with a precision of ~0.2dex compared with high-resolution analysis. The recommended FCA threshold value {lambda} for this test is between 0.9965 and 0.9975. After reducing the dimension of the line indices through FCA, the derived metallicities are still robust, with no loss of accuracy, and the FCA-ANN method performs stably for different spectral quality from [Fe/H]~1.8 down to -3.5. Compared with traditional classification methods, FCA considers ambiguity in groupings and noncontinuity of data, and is thus more suitable for observational data analysis. Though this early test uses FCA to analyze low-resolution spectra, and feeds the input to the ANN method to derive metallicities, FCA should be able to, in the large data era, also analyze slitless spectroscopy and multiband photometry, and prepare the input for methods not limited to ANN, in the field of stellar physics for other studies, e.g., stellar classification, identification of peculiar objects. The literature-collected high-resolution sample can help improve pipelines to derive stellar metallicities, and systematic offsets in metallicities for VMP stars for three published LAMOST catalogs have been discussed.