I suspect it works on many different levels and that the general stress everyone feels from it is one of those levels.

But in terms of the base "object level" of the actual graphing of multiple biometrics, the theoretical idea is to construct a baseline within the confines of the test, not between you inside and outside of the test. They let you settle in, ask you questions they know the answer to, and compare measurements during known truthful responses to measurements during or near other responses. So they're (theoretically) accounting for the added stress in the baseline.

Obviously this object-level portion of detecting honesty vs. deceit via comparison with a baseline of physiological characteristics is pseudoscientific and extremely unreliable, but it's part of the overall mind game. And even though it's pseudoscientific, it's also not like it's palm reading or something. This sounds a bit ridiculous to say, but pseudoscience is a spectrum. (For example, astrology is more pseudoscientific than MBTI.) If someone has totally normal responses to every question but a bizarrely elevated response to one single seemingly harmless question which no one else has an unusual response to, then that's something to look into and prod. So the object level of it can offer some utility.