It depends on the particular SDE. There is no systematic approach.
One main idea is striving for some cancelation to simplify the drift/volatility coefficients (eg.https://math.stackexchange.com/questions/823785/stochastic-differential-equation-with-trigonometric-functions). For example, in Geometric Brownian motion
$$ dS_t = \mu S_t\,dt + \sigma S_t\,dW_t$$
one observes the $S_{t}$ term would cancel out with the derivative of $f(x)=log(x)$.
Another idea is to help transfer some ODE techniques such as integration-factor see here.
Solution to General Linear SDE