Robust screening of atrial fibrillation with distribution classification

Figures

• 

  Fig. 1. The summary of our method, an SVM for distribution classification. It classifies a new, unseen patient by comparing the full empirical distribution of its RRis to those of the training set, enhancing robustness and adaptability to inputs of varying sizes.

• 

  Fig. 2. Cross-data-set ROC curves.

• 

  Fig. 3. Confusion matrices of the classifier trained on different data sets and evaluated on the SPH data set. True positives and negatives are circled in gray. Acronyms on the y-axis are those defined in the documentation of the SPH data set26, with the exception of atrial flutter and atrial fibrillation, which we rename as “AFlut” and “AF” instead of “AF” and “AFib”, respectively. In short, AF is atrial fibrillation, SR is normal sinus rhythm, and other labels are conditions other than AF. We see that training cross-data-set on the DiagnoStick or CinC data set only marginally changes the matrix, indicating that most examples are classified identically and showing robustness. Further, normal sinus rhythm is almost perfectly separated from AF.

• 

  Fig. 4. Testing performance with increasing training set size. Each box plot is obtained with 100 independent resamplings of the training set.