Typical feature selection methods choose an optimal global feature subset that is applied over all regions of the sample space. In contrast, in this talk we propose the novel localized feature selection (LFS) approach whereby each region of the sample space is associated with its own distinct optimal feature subset, which may vary both in membership and size across the sample space. This allows the feature set to optimally adapt to local variations in the sample space. An associated method for measuring the similarities of a query datum to each of the respective classes is also proposed. The team demonstrate the method is robust against the over-fitting problem and that it is effective when the number of candidate features far exceeds the number of training samples.

The team apply the LFS method to the problem of predicting emergence of coma patients. Previous work has shown that the presence of the mismatch negativity (MMN) component of the event-related potential (with respect to the EEG) has strong correlation with emergence. The difficulty with current approaches is that the MMN is very difficult to detect, rendering low sensitivity values. The result show some clinical evidence that suggests the LFS method is useful in this context.