Hi,
1. I would use the same number of subjects/training samples for each group so that chance accuracy is the same across groups. However, you can also correct for bias if you are using an unequal number of training samples across groups.
2. The -predictions option will give you the distance to the hyperplane for each binary classifier <prefix>_<class label A>_<class label B>.1D and overall muticlass class predictions (<prefix>_DAG.1D or <prefix>_vote.1D). Given this output and and the labels, you can calculate sensitivity and specificity for each class. We currently only report accuracy to standard out.
Please note that the default output (binary classifiers) for -predictions is centered around 0.5 (hyperplane at 0.5), so everything > 0.5 is "class label A" and everything <= 0.5 is "class label B", where "class label A" > "class label B". If you want the predictions to be centered around 0, please use flag: -nopredscale. Please also note that 3dsvm was originally written for classification of temporal data and that temporal detrending of the classifier output is enabled by default. Use flag: -nodetrend to disable detreding.
3. I would use the -censor option for this. This allows you to include/exclude subjects without having to create a new dataset for each randomization/cross-validation step.
Jonathan