A novel method for extraction and classification of ischemic features from electrocardiograms, based on the dilated discrete Hermite expansion, is described. The discrete Hermite functions used for the expansion are eigenvectors of a symmetric tridiagonal matrix that commutes with the centered Fourier matrix. A choice of 50 Hermite coefficients and a dilation parameter were sufficient to reconstruct the ECG with all essential features preserved. The performance was measured using percentage RMS difference (PRD). The 50 coefficients and the dilation parameter contain information about the shape of the ECG and a committee neural network classifier with these 51 input parameters was trained to identify ischemic features. A sensitivity of 97% and a specificity of 94% was achieved. This technique can also be used for training neural networks to identify other abnormalities of the ECG.