Boolean networks offer an elegant way to model the behaviour of complex systems with positive and negative feedback. The long-term behaviour of a Boolean network is characterised by its attractors. Depending on various logical parameters, a Boolean network can exhibit vastly different types of behaviour. Hence, the structure and quality of attractors can undergo a significant change known in systems theory as attractor bifurcation. In this paper, we establish formally the notion of attractor bifurcation for Boolean networks. We propose a semi-symbolic approach to attractor bifurcation analysis based on a parallel algorithm. We use machine-learning techniques to construct a compact, human-readable, representation of the bifurcation analysis results. We demonstrate the method on a set of highly parametrised Boolean networks.