We propose a novel framework to predict the location of a myocardial infarct from local wall deformation data. Non-linear dimensionality reduction is used to estimate the Euclidean space of coordinates encoding deformation patterns. The infarct location of a new subject is inferred by two consecutive interpolations, formulated as multiscale kernel regressions. They consist in (i) finding the low-dimensional coordinates associated to the measured deformation pattern, and (ii) estimating the possible infarct location associated to these coordinates. These concepts were tested on a database of 500 synthetic cases generated from a realistic electromechanical model of the two ventricles. The database consisted of infarcts of random extent, shape, and location overlapping the whole left-anterior-descending coronary territory. We demonstrate that our method is accurate and significantly overcomes the limitations of the clinically-used thresholding of the deformation patterns (average area under the ROC curve of 0.992±0.011 vs. 0.812±0.124, p<0.001).