It's simply a Fourier transform relationship. If an atom is in a metastable state, with the latter coupled to the electromagnetic field, then it is fairly straightforward to show that the probability amplitude for decay to happen in the time interval $[t,\,t+\mathrm{d}t]$ is:
$$\psi(t) = \left\{\begin{array}{ll}\frac{1}{\sqrt{\tau}}\,\exp\left(i\,\omega_0\,t - \frac{t}{2\,\tau}\right);& t\geq0\\0&\text{otherwise}\end{array}\right.$$
where $\hbar\,\omega$ is the transition energy. I show how this calculation is done in my answer here. Witness that the probability density $|\psi|^2$ is the memoryless exponential distribution with lifetime $\tau$. If you work out the Fourier transform of $\psi$ then the power spectral density to get the lineshape, you get a Lorentzian lineshape whose linewidth is given by the relationship you ask about.
The Heisenberg-like inequalities that people pull out of thin air without explanation are all manifestations of the relationships between second moments of functions and their power spectral densities. See my treatment of this question here. If you don't want to spend your life feeling like killing the next person who shoves slapdash Heisenberg like relationships down your throat without explanation as though it were some kind of Physicist Masonic handshake, then you need to study these kinds of behaviors of the Fourier transforms carefully. Even the Heisenberg uncertainty principle itself ultimately comes down to this (or this is one way to derive it): quantum states in conjugate co-ordinates are mapped to one-another by the Fourier transform.
