Detection of Ventricular Fibrillation Based on Ballistocardiography by Constructing an Effective Feature Set

Figures

• 

  Figure 1. Flow diagram of the proposed method.

• 

  Figure 2. Process of the animal experiment.

• 

  Figure 3. Comparison before and after adaptive filtering. (b,e) represent BCG signals after adaptive filtering in (a,d), respectively, where the arrows in (a,d) point to the pulsed artifact interference caused by the ventilator. (c,f) represent residuals between the raw BCG and the filtered BCG, namely the interference components that are filtered out.

• 

  Figure 4. Time-domain BCG and ACF sequences. (a–d) belongs to VF, (e–h) belongs to SR, and (i–l) belongs to MA; where (b,d,f,h,j,l) represent the ACF sequences of time-domain BCG signals (a,c,e,g,i,k), respectively.

• 

  Figure 5. Function visualization results in the process of feature extraction. (a–d), VF; (e–h), SR; (i–l), MA, where (a,e,i) are original time-domain signals Tt; (b,f,j) are two-dimensional time-frequency signals Sf,t; (c,g,k) are one-dimensional signals S¯t obtained by averaging along the frequency axis; (d,h,l) are one-dimensional signals S¯f obtained by averaging along the time axis.

• 

  Figure 6. Mean and standard deviation (SD) of each of the features (in Table 2) in cases of VF, SR, and MA. The y-axis scale is normalized against the global maximum value of each specific feature (in VF, SR, and MA classes jointly).

• 

  Figure 7. Hierarchical clustering heatmap of features. According to the red dividing line on the clustering tree, 12 classes can be aggregated, and the results are circled in red boxes.

• 

  Figure 8. Example of misjudged signals. (a–c), VF; (d,e), SR; (f), MA. The BCG time-domain waveform and the corresponding ACF sequence are shown in blue and orange, respectively.