Paroxysmal Atrial Fibrillation (PAF) may lead to the decline of atrial mechanism and hemodynamic disturbance, which is the major risk factor for stroke, myocardial infarction, heart failure and other diseases. Although a number of methods have been developed to detect PAF, they rely on long ECG signals. Furthermore, most studies are only for the intra-patient test, ignoring individual specificity. Therefore, this paper proposes an automatic detection algorithm based on the combination of RR interval time-domain and nonlinear features to detect short-term PAF. The proposed features reflect the underlying physiological phenomenon of highly irregular ventricular activity caused by atrial unstable conduction. These include changes in morphological features and intrinsic changes in the cardiac dynamic system, which can capture subtle changes in early short-term PAF episodes and attenuate specificity between individuals. Finally, the random forest algorithm was used to detect the MIT-BIH Atrial Fibrillation Database and MIT-BIH Arrhythmia Database with intra-patient and inter-patient paradigm. The results show that the two databases’ algorithm accuracy in different paradigms of intra-patient and inter-patient have achieved more than 99.50% and 95.50%, respectively. Therefore, the proposed algorithm has a potential application for clinical monitoring.