I had the same problem in Rapahel 2.2.0 and 2.2.1, and using ._getBBox() didn't fix it for me. 
What did fix it for me is falling back to .auxGetBBox() if it's defined and regular .getBBox() doesn't work, like this:
var bbox = path.getBBox( );
// Workaround for apparent bug in Raphael's VML module's getBBox() override
if( isNaN( bbox.x ) && path.auxGetBBox ){
  bbox = path.auxGetBBox();
}
I don't have a fix for the underlying bug, but I have found the source of it.
In VML mode, Raphael takes the initial getBBox() function, saves it as auxGetBBox() on the element prototype, then replaces it with a function that appears to be broken. 
It has calculations based on a variable defined as var z = 1/this.paper._viewBoxShift.scale;, which clearly expects _viewBoxShift.scale to be some factor of the scale of the current viewbox compared to the initial viewbox , but actually _viewBoxShift.scale is an object like this which appears to come from paperproto.getSize():
{ height: someNumber, width: someNumber }
This is where all the NaNs are coming from. Cannae divide by an object.
So this workaround works fine if no zoom is applied using a viewbox, but may give incorrect results if a zoom has been applied (something I can't get to work at all in recent versions of raphael in VML mode, but that's a seperate question). Fixing that will involve digging deep into Raphael's VML module to pipe a proper zoom factor into this z variable.