Atrial fibrillation is the most common cardiac arrhythmia in clinical practice, while some realtime automatic detection algorithms have been developed rapidly for improving clinical diagnosis. However, it is difficult to judge the advantages and disadvantages of the existing atrial fibrillation detection algorithms for their lack of comparisons between models or databases. Three different machine learning algorithms, including support vector machine, random forest and logistic regression, were selected to build three separate models to detect atrial fibrillation. The three models were trained on the MITBIH atrial fibrillation database, and were tested and compared on three independent databases respectively. Furthermore, the influence of feature selection on model performance was analyzed. Experimental results showed that when applied 12 features (3 domain features and 9 nonlinear features), the sensitivity, specificity, accuracy and F1score of the three models reached more than 95% on the China physiological signal challenge 2018 public database and the wearable dynamic ECG database. In addition, the random forest algorithm has stronger stability and generalization ability compared with the other two algorithms.