Using Minimum Redundancy Maximum Relevance Algorithm to Select Minimal Sets of Heart Rate Variability Parameters for Atrial Fibrillation Detection

Figures

• 

  Figure 1. Area under receiver operating characteristic curve (AUC) of heart rate variability features. pRR50—percentage of successive differences between RR intervals greater or equal to 50 ms; SD1 and SD2—standard deviation of points in the Poincare plot across and along the identity line, respectively; SDRR—standard deviation of RR intervals; RRdif—mean of absolute differences between successive RR; CV—coefficient of variance; relRRdif = RRdif/(mean RR); mean RR—mean of RR intervals; RRrange = max(RR) − min(RR); relRRrange = RRrange/(mean RR) (relative RRrange).

• 

  Figure 2. Ranking of heart rate variability features by minimum redundancy maximum relevance (MRMR) algorithm. pRR50—percentage of successive differences between RR intervals greater or equal to 50 ms; SD1 and SD2—standard deviation of points in the Poincare plot across and along the identity line, respectively; mean RR—mean of RR intervals; SDRR—standard deviation of RR intervals; RRdif—mean of absolute differences between successive RR; CV—coefficient of variance; relRRdif = RRdif/(mean RR); meanSuccRat—mean ratio of successive RR; SDSuccRat—standard deviation of ratios of successive RR; RRrange = max(RR) − min(RR); relRRrange = RRrange/(mean RR) (relative RRrange).

• 

  Figure 3. Histograms of six heart rate variability parameters with the highest score from minimum redundancy maximum relevance algorithm (blue—sinus rhythm, SR; orange—atrial fibrillation, AF). pRR50—percentage of successive differences between RR intervals greater or equal to 50 ms; SD1 and SD2—standard deviation of points in the Poincare plot across and along the identity line, respectively; CV—coefficient of variance; mean RR—mean of RR intervals; relRRdif—mean of absolute differences between successive RR divided by mean RR; relRRrange = (max(RR) − min(RR))/(mean RR) (relative RRrange).

• 

  Figure 4. Diagram of training and validation methodology. MRMR—minimum redundancy maximum relevance.

• 

  Figure 5. Accuracy of AF detection for different classifiers with sets of one to six features (orange points—blindfold validation, blue points—5-fold cross-validation). DT—decision tree, KNN—K nearest neighbors, SVM linear—support vector machine with linear kernel, SVM RBF—support vector machine with radial basis function kernel, ADA—Ada Boost, RF—random forest, ANN—artificial neural network.

• 

  Figure 6. Sensitivity of atrial fibrillation detection for different classifiers with sets of one to six features (orange points—blindfold validation, blue points—5-fold cross-validation). DT—decision tree, KNN—K nearest neighbors, SVM linear—support vector machine with linear kernel, SVM RBF—support vector machine with radial basis function kernel, ADA—Ada Boost, RF—random forest, ANN—artificial neural network.

• 

  Figure 7. Specificity of AF detection for different classifiers with sets of one to six features (orange points—blindfold validation, blue points—5-fold cross-validation). DT—decision tree, KNN—K nearest neighbors, SVM linear—support vector machine with linear kernel, SVM RBF—support vector machine with radial basis function kernel, ADA—Ada Boost, RF—random forest, ANN—artificial neural network.

• 

  Figure 8. Positive predictive value (PPV) of AF detection for different classifiers with sets of one to six features (orange points—blindfold validation, blue points—5-fold cross-validation). DT—decision tree, KNN—K nearest neighbors, SVM linear—support vector machine with linear kernel, SVM RBF—support vector machine with radial basis function kernel, ADA—Ada Boost, RF—random forest, ANN—artificial neural network.

• 

  Figure 9. The diagnostic odds ratio (DOR) of AF detection for different classifiers with sets of one to six features (orange points—blindfold validation, blue points—5-fold cross-validation). DT—decision tree, KNN—K nearest neighbors, SVM linear—support vector machine with linear kernel, SVM RBF—support vector machine with radial basis function kernel, ADA—Ada Boost, RF—random forest, ANN—artificial neural network.

• 

  Figure 10. Ranking of heart rate variability features by minimum redundancy maximum relevance (MRMR) algorithm after excluding percentage of successive differences between RR intervals greater or equal to 50 ms (pRR50). SD1 and SD2—standard deviation of points in the Poincare plot across and along the identity line, respectively; mean RR—mean of RR intervals; SDRR—standard deviation of RR intervals; RRdif—mean of absolute differences between successive RR; CV—coefficient of variance; relRRdif = RRdif/(mean RR); meanSuccRat—mean ratio of successive RR; SDSuccRat—standard deviation of ratios of successive RR; RRrange = max(RR) − min(RR); relRRrange = RRrange/(mean RR) (relative RRrange).

• 

  Figure 11. Accuracy of atrial fibrillation detection for different classifiers with sets of one to six features without pRR50 (orange points—blindfold validation, blue points—5-fold cross-validation). DT—decision tree, KNN—K nearest neighbors, SVM linear—support vector machine with linear kernel, SVM RBF—support vector machine with radial basis function kernel, ADA—Ada Boost, RF—random forest, ANN—artificial neural network.

• 

  Figure 12. Diagnostic odds ratio (DOR) of atrial fibrillation detection for different classifiers with sets of one to six features (orange points—blindfold validation, blue points—5-fold cross-validation). DT—decision tree, KNN—K nearest neighbors, SVM linear—support vector machine with linear kernel, SVM RBF—support vector machine with radial basis function kernel, ADA—Ada Boost, RF—random forest, ANN—artificial neural network.