Objective To propose an electrocardiogram(ECG) feature extraction method,which is able to reflect the difference of the importance existing among the classes and at the same time to overcome the deficiency of conventional feature extraction method that is unable to solve the classification problem with priority.Methods The data for analysis was obtained from MIT-BIH database,including 250 samples each of normal sinus rhythm(NSR),atria premature contraction(APC),premature ventricular contraction(PVC),ventricular tachycardia(VT),ventricular fibrillation(VF) and super-ventricular tachycardia(SVT).The projecting vectors were constructed following the introduction of separable measurement to extract the features for the classification with priority.Results The proposed feature extraction method increased the average accuracy by 12 percentages as compared with conventional linear discriminative analysis(LDA) methods.Conclusion The prior classes can be better discriminated from others and separate each of the classes as much as possible at the same time.