It is important for cardiac therapy devices such as the automated external defibrillator to discriminate different cardiac disorders based on Electrocardiogram analysis. A phase space analysis based algorithm is proposed to detect cardiac arrhythmias effectively. Firstly, the phase space of the signal is reconstructed. Then from the viewpoint of geometry and information theory, the distribution entropy of the point density in the two-dimensional reconstructed phase space is calculated as the features in the further classification. Finally the nearest-neighbour method based on Mahalanobis distance is used to classify the sinus rhythm (SR), supraventricular tachyarrhythmia (SVTA), atrial flutter (AFL) and atrial fibrillation (AF). To evaluate the sensitivity, specificity and accuracy of this proposed method in the cardiac arrhythmias classification, the MIT-BIH arrhythmias database and the canine endocardial database are studied respectively. Experiment results demonstrate that the proposed method can detect SR, SVTA, AFL and AF signals rapidly and accurately with the simple computation. It promises to find application in automated devices for cardiac arrhythmias therapy.