A standard Neurofeedback approach to mitigate the working memory decline in some fragile groups (elderly, subjects affected by stroke or Alzheimer's disease) can be suboptimal for some patients. The goal of this research is to investigate which visual stimulus (among colour, geometrical shape, direction, and symbol) is the most suited for each of the six healthy participants and which brain areas are the most discriminative, during the maintenance of a presented stimulus in a retro-cue-based working memory experiment. In order to identify the most discriminative stimulus, the single-trial classification accuracies of some Support Vector Machines, trained on the theta, alpha and beta electroencephalography power bands, have been compared; while, in order to identify the most involved brain regions, three machine learning feature reduction techniques have been explored: the first based on a massive univariate analysis, the second based on multivariate filtering and wrapping, and the last one based on Frequency-based Common Spatial Pattern. The results have shown that the univariate approach, more than the others, managed to clearly identify for each participant at least one preferential type of stimulus and a brain region of discriminative electrodes during the maintenance of the stimulus. These promising results can be interpreted as a further step to optimize the Neurofeedback working memory enhancement through a personalised approach.