Dimension reduction is essential in fMRI decoding, but the complex relationship between fMRI data and class labels is often unknown or not well modeled so that the most effective dimension reduction (e.d.r.) directions can hardly be identified. In the present study, we introduce a novel fMRI decoding approach based on an effective and general dimension reduction method, namely sliced inverse regression (SIR), which can exploit class information for estimating e.d.r. directions even when the relationship between fMRI data and class labels is not explicitly known. We incorporate singular value decomposition (SVD) into SIR to overcome SIR’s limitation in dealing with ultra-high-dimensional data, and integrate SVD-SIR into a pattern classifier to enable quantification of the contributions of fMRI voxels to class labels. The resultant new SIR decoding analysis (SIR-DA) approach is capable of decoding behavioral responses and identifying predictive fMRI patterns. Simulation results showed that SIR-DA can more accurately detect e.d.r. directions and achieve higher classification accuracy than decoding approaches based on conventional dimension reduction methods. Further, we applied SIR-DA on real-world pain-evoked fMRI data to decode the level of pain perception and showed that SIR-DA can achieve higher accuracy in pain prediction than conventional methods. These results suggest that SIR-DA is an effective data-driven technique to decode behavioral or cognitive states from fMRI data and to uncover unknown brain patterns associated with behavior or cognitive responses.