Yes, you can use "fsrftest" for feature selection before training a "fitrensemble" model with LS-Boost, but it is not always necessary.
Tree-based models like LS-Boost already perform feature selection during training by choosing informative splits, so they naturally down-weight unimportant features. Applying "fsrftest" beforehand might help rus to educe dimensionality if we are having hundreds or thousands of features, but there is a risk of discarding useful features, if they interact in ways the filter does not capture.
Since, the results will be depend on the dataset, you can apply "fitrensemble" with and without "fsrftest" and evaluate the performance.
