Detection of Atrial Fibrillation Using Decision Tree Ensemble

2017 PhysioNet/CinC Challenge proposed a global competition for classifying a short single ECG lead recording into normal sinus rhythm, atrial fibrillation (AF), alternative rhythm, and unclassified rhythm. This study developed and evaluated a pragmatic approach to solve the challenge, which was based on a decision tree ensemble with 30 features from ECG recording. The model was trained using the AdaBoost.M2 algorithm. The results reported here were obtained using 100-fold cross-validation, and the lowest MSE was 0.12 with the maximum number of splits of 55, and the number of trees of 20. The entry was tested and scored in the second phase of the challenge. The achieved scores for "Normal", "AF", "Other", were 0.93, 0.86, and 0.79, respectively, while the F1 measure was 0.86, and the official overall score was 0.82.