The increasing cardiac diseases of people in recent years demand an early detection of heart diseases using electrocardiogram (ECG) signal processing techniques. In this work we present a semi automatic scheme to discriminate patient-specific ECG beats by using a kernel based feature extraction technique called kernel canonical correlation analysis (KCCA). The heartbeat classification scheme uses a fixed global training set along with a local patient specific training set to produce a patient adapted training set. The performance of the KCCA features is verified on MIT-BIH arrhythmia database by using three classifiers, namely k-nearest neighbor (KNN), support vector machine (SVM) and vector valued regularized kernel function approximation (VVRKFA), to classify four different type of beats, normal (N), supra-ventricular (S), ventricular (V) and fusion of ventricular and normal (F). Experimental results show that KCCA features are more effective on normal patients than that of arrhythmia patients.