Arrhythmia is an abnormal beat result from any disorder in the conduction of the cardiac electrical impulse. It is very essential to identify and detect arrhythmias correctly in its early phases. Manual diagnosis of arrhythmia beats is very difficult due to their unfamiliar mechanisms and composite nature. Current paper introduces, a machine learning-based methodology proposed for automated cardiac arrhythmia detection. The methodology follows ECG filtering and segmentation using general approach, followed by Discrete Cosine Transform (DCT) for feature extraction, Principal Component Analysis (PCA) for feature reduction and finally classification of arrhythmia beats using k-Nearest Neighbor (k-NN) classifier. In this study, the statistical significance of PCA features is verified using Analysis of Variance (ANOVA) test. Statistically significant features are classified using k-NN and tenfold cross validation. In this study, all the beats of entire MIT-BIH (Massachusetts Institute of Technology—Boston’s Beth Israel Hospital) arrhythmia database are considered. The five classes recommended by ANSI/AAMI EC57:1998 standard: Non-ectopic (N), Supraventricular ectopic (S), Ventricular ectopic (V), Fusion (F) and Unknown (U) cardiac beats are recognized with an average class specific accuracy of 99.93, 98.41, 98.09, 96.93, 99.7 % respectively and overall average accuracy of 98.61 %. Added validation of the proposed method can result in suitable outcome for therapeutic applications.