I used this in my PhD thesis in 1995 almost as a folk result: it is elementary spectral theory. I recall reading, on the probabilistic side, papers related to the subject by P.A. Meyer from the '70s, and did not claim any originality, but rather pointed to the book on Littlewood-Paley theory by Stein. In short, after taking a Fourier transform the Laplace equation $\partial_{tt}u+\Delta_xu=0$ in ${\mathbb R}^{n+1}_+\ni(x,t)$ space becomes $\partial_{tt}\hat{u}(\xi,t)-|\xi|^2\hat{u}(\xi,t)$, whose solutions are $$ \hat{u}(\xi,t)=e^{-t|\xi|}\hat{f}(\xi)=0, $$ or $$ u(x,t)=e^{-t\sqrt{-\Delta_x}}f. $$ The interpretation Dirichlet to Neumann immediately follows. Given the kind of argument, I would be very surprised if this was not commonplace already at the beginning of the 20th century, or even before.
As you can take the square root of any positive operator, this simple reasoning extends to a variety of situations. P.A. Meyer took as positive operators those coming from the "carre' du champ", which extend the infinitesimal generators of stochastic process, which include second order elliptic operators...
I will look for a reference, if this is still of interest for you.