Functional magnetic resonance imaging (fMRI) data analysis is a challenging problem due to the underlying physiological complexity of the brain and the scanning process. From engineering perspective, the fMRI data analysis can be viewed as a system modelling problem. In this paper, assuming the fMRI signal as the output of an unknown linear time-invariant system, a spatiotemporal adaptive filter is proposed to model the spatial activation patterns as well as the haemodynamic response (HDR) to the event-related stimulus. The well-known least mean square adaptive algorithm is used for estimating the coefficients of the spatiotemporal filter. The proposed method is shown to be equivalent to the canonical correlation analysis method. It is then extended to multiple event type scenarios to estimate the HDRs of each event type. Results from simulated as well as real fMRI data show that these adaptive modelling schemes can capture the variations of the HDR at different regions of the brain and hence enhance the estimation accuracy of the activation patterns.