Within a Heartbeat: A Machine Learning Approach Using VCG for Heart Failure Assessment in Sinus Rhythm

Figures

• 

  Figure A1. Denoising process step by step on a PTB signal (left) and a MUSIC signal (right). (a) raw vs. detrended signal (PTB); (b) raw vs. trimmed signal (MUSIC); (c) detrended vs. baseline removed signal (PTB); (d) trimmed vs. detrended signal (MUSIC); (e) baselined removed vs. spike-corrected signal (PTB); (f) detrended signal vs. baseline removed signal (MUSIC); (g) spike-corrected vs. notch filtered signal (PTB); (h) baselined removed vs. spike-corrected signal (MUSIC); (i) notch filtered vs. wavelet denoised signal (PTB); (j) spike-corrected vs. notch filtered signal (MUSIC); (k) wavelet denoised vs. bandpass filtered signal (PTB); (l) notch filtered vs. wavelet denoised signal (MUSIC); (m) wavelet denoised vs. bandpass filtered (MUSIC).

• 

  Figure A2. Comparison of denoising pipeline on an original raw signal and a copy of it with added artificial noise—before denoising. (a) Whole signal; (b) Segment of signal (25–30 s).

• 

  Figure A3. Comparison of denoising pipeline on an original raw signal and a copy of it with added artificial noise—after denoising. (a) Whole signal; (b) Segment of signal (25–30 s).

• 

  Figure A4. Outlier detection evaluation on a clean signal. (a) Visualization of R peaks (outliers not found) across the whole signal; (b) All extracted heartbeats plotted together—each gray line represents each heartbeat and the blue line the average heartbeat; (c) Scatter matrix of PC1, PC2 and PC3. Diagonal panels show the distribution (histogram) of each PC and off-diagonal panels show pairwise scatter plots (outliers were not found).

• 

  Figure A5. Outlier detection evaluation on a contaminated signal. (a) Visualization of valid R peaks and outliers across the whole signal; (b) All extracted heartbeats plotted together—each gray line represents each heartbeat, each red line an outlier beat, and the blue line the average heartbeat; (c) Scatter matrix of PC1, PC2 and PC3. Diagonal panels show the distribution (histogram) of each PC and off-diagonal panels show pairwise scatter plots with outliers marked in red.

• 

  Figure A6. Batch effect analysis including healthy subjects from PTB and HF patients from MUSIC and PTB—PC1 + PC2.

• 

  Figure A7. Batch effect analysis including healthy subjects from PTB and HF patients from MUSIC and PTB—PC1 + PC2 + PC3.

• 

  Figure A8. Waterfall SHAP plot—HFpEF patient 1. Red color indicates features that push the prediction toward the positive class (HFpEF).

• 

  Figure A9. Waterfall SHAP plot—HFpEF patient 2. Red color indicates features that push the prediction toward the positive class (HFpEF) while blue color indicates features that push the prediction toward the negative class (normal).

• 

  Figure A10. Waterfall SHAP plot—HFrEF patient 1. Red color indicates features that push the prediction toward the positive class (HFrEF) while blue color indicates features that push the prediction toward the negative class (normal).

• 

  Figure A11. Waterfall SHAP plot—HFrEF patient 2. Red color indicates features that push the prediction toward the positive class (HFrEF) while blue color indicates features that push the prediction toward the negative class (normal).

• 

  Figure 1. Feature distribution with raw and corrected p-values (FDR) from the KW test—normal vs. HFpEF dataset; the orange line represents the median while the circles represent outliers.