On 2015 Feb 25, Gerard Ridgway commented:

This work is generally very thorough, but the investigation of discriminant analysis seems rather limited. Only linear and quadratic discriminant analysis are considered, and there is little detail about the methods (for example, no mention of dimensionality reduction or regularisation terms). With up to 180 features, regularised and/or kernel variants of LDA, or other methods such as support vector machines (used in e.g. Ecker C, 2010) might well perform considerably better.

It could be an interesting exercise if the authors would be willing to share their features (to save other researchers from the time-consuming process of running FreeSurfer on many subjects) along with labels for a randomly selected half of the subjects. Other researchers could then submit their predictions for the unlabelled half to Haar et al. for evaluation of the accuracy. I believe the numbers here would be sufficient for a split-half validation to provide meaningful results.

On 2015 Feb 25, Gerard Ridgway commented:

This work is generally very thorough, but the investigation of discriminant analysis seems rather limited. Only linear and quadratic discriminant analysis are considered, and there is little detail about the methods (for example, no mention of dimensionality reduction or regularisation terms). With up to 180 features, regularised and/or kernel variants of LDA, or other methods such as support vector machines (used in e.g. Ecker C, 2010) might well perform considerably better.

It could be an interesting exercise if the authors would be willing to share their features (to save other researchers from the time-consuming process of running FreeSurfer on many subjects) along with labels for a randomly selected half of the subjects. Other researchers could then submit their predictions for the unlabelled half to Haar et al. for evaluation of the accuracy. I believe the numbers here would be sufficient for a split-half validation to provide meaningful results.