Myocardial deformation is a critical indicator of many cardiac diseases and dysfunctions. The goal of this paper is to use myocardial deformation patterns to identify and localize regional abnormal cardiac function in human subjects. We have developed a novel tensor-based classification framework that better conserves the spatio-temporal structure of the myocardial deformation pattern than conventional vector-based algorithms. In addition, the tensor-based projection function keeps more of the information of the original feature space, so that abnormal tensors in the subspace can be back-projected to reveal the regional cardiac abnormality in a more physically meaningful way. We have tested our novel method on 41 human image sequences, and achieved a classification rate of 87.80%. The recovered regional abnormalities from our algorithm agree well with the patient's pathology and doctor's diagnosis and provide a promising avenue for regional cardiac function analysis.