Inverse solution allows to estimate sources that generate a given scalp EEG topography. Recently, it has been used in Brain Computer Interfaces (BCIs) to extract robust features based on the hypothesis that projection onto the source space (high dimensional space) yield promising improvement in classification performance over surface EEG through offline experiments. However, these inverse methods are optimized for localizing the intracranial sources of the EEG signals. Therefore, using inverse solution in its current form may not be optimal for classification purposes and its online application is yet to be proven. To this end, this thesis aims at characterizing different parameters associated with inverse solution such that it results in increased classification performance when applied online with respect to surface EEG while maintaining its source localization capabilities. This thesis documents three major contributions : (i) it presents methods for feature selection and classification using inverse solution and surface EEG for Event Related Potential (ERP) based BCIs, (ii) it reports on the optimal choice of parameters for inverse solution to balance between localization and classification for BCI using the developed feature selection and classification methods, (iii) it applies this procedure on four BCI experiments and shows that it improves classification performance over surface EEG in online conditions. First, we focus on developing feature selection and classification methods that are capable to run for online ERP-based BCIs. The use of inverse solution results in an increase in the available number of features. To select features, we apply Fisher score to quantify the discriminability of each source and select the same number of top-most discriminant sources at each time instance in a selected time window. To classify, an ensembled approach was followed where we made individual classifier for each feature and combined them using Naïve Bayes rule to obtain the resultant decision. This procedure prevents overfitting and requires only the number of sources to be optimized. For comparison with surface EEG, we applied Fisher Linear Discriminant (FLD) spatial filters at each time instance on the potential from all the EEG channels. Individual classifiers are trained on each FLD projection and are combined using the same rule. This comparison is extended with two additional classifiers for surface EEG: multiple projection FLD and specific channel classifier. Second, the parameters of the inverse solution are characterized in terms of their localization and classification capabilities for ERP based BCI. The parameters include forward head model, regularization constraint and regularization parameter. We find that, the choice of head model does not affect the optimum classification performance (although, it clearly affects the localization process) that can be achieved with our feature selection and classification methods. For a chosen forward model, the values of the regularization parameter that yield optimum classification performance remain consistent across subjects and experiments. The range however depends on regularization constraint. Furthermore, we find that the same range results in minimum localization error for different noise levels in EEG. Regarding standardized regularization constraint, although they can achieve lower localization error, they also have lower classification performance. Finally, the proposed feature selection and classification methods are applied on four different ERP experiments. The first three experiments: ErrP, RSVP and P300-speller were performed offline and were used to characterize the number of features for inverse solution that exhibit optimum classification performance. We found that the classification performance with a small number of features was higher compared to surface EEG and overall is less sensitive to the number of features. The results for classification performance using inverse solution are in-line with the literature on ERP based BCI. Following the analysis on offline experiments, we showed the feasibility of using inverse solution to online ErrP experiment. In summary, this thesis provides an insight on the use of estimated intracranial sources as features for online BCI applications. Specifically, it characterizes parameter choice along with feature selection and classification methods. This results in high classification performance while retaining low localization error. Overall, the methods developed are easily applicable to ERP based BCIs as demonstrated in both offline and online experiments.