Heartbeat Detection by Laser Doppler Vibrometry and Machine Learning

Figures

• 

  Figure 1. The LDV measurement set-up is composed by a Laser Doppler Vibrometer (LDV) pointed on the subject’s skin over the carotid. The LDV signal is generated and fed to a PC by a DAQ board.

• 

  Figure 2. Workflow of the proposed machine-learning (ML) framework to heartbeat detection from carotid LDV signal.

• 

  Figure 3. Sliding-window algorithm used to extract LDV signal windows. Wd window size, Ws =shift length, Woverlap = overlap with a new beat. The green and blue windows denote beat and no-beat samples, respectively. (The LDV signal was smoothed for visualization purposes only).

• 

  Figure 4. Each curve refers to one subject and is the average of a portion of the LDV signal. The portion starts from the R peak in the electrocardiograph (ECG) and last 800 ms. The red circle shows the LDV pulse wave peak that can be found at 102 ± 15 ms after the R-peak.

• 

  Figure 5. Receiver operating characteristic (ROC) curves obtained using the LDV-beat dataset with (a) Random Forest (RF), (b) Decision Tree (DT), (c), K-Nearest Neighbor (KNN) and (d) Support Vector Machines (SVM). Each ROC refers to a subject. The area (AUC) under the ROC is shown, too.

• 

  Figure 6. Each curve refers to one subject and is the average of a portion of the LDV signal. The portion starts from the R peak in the ECG and last 800 ms. The red curves refer to the subjects with the lowest AUC values.

• 

  Figure 7. Classification results obtained on the LDV-beat dataset. Results are shown in terms of: (a) recal (Reci), (b) precision (Preci), and (c) f1-score (f1i) for the class i = beat. The results are obtained when using Random Forest (RF), Decision Tree (DT), K-Nearest Neighbor and Support Vector Machines (SVM).

• 

  Figure 8. Four sample testing results on the full-length (10-min long), LDV signal (the signal was smoothed for visualization purposes only). In each subfigure, the ECG signal with gold-standard beats (grey rectangles centered with the R peak from the ECG), and the output probability (vertical blue line) for the beat class obtained with support vector machines are shown. The beats in the ECG signal are shown with rectangles for better visibility.

• 

  Figure 9. Classification results in terms of macro-average metrics obtained when classifying the full length (10-min long) LDV signals using a Support Vector Machine (SVM). Results are shown in terms of: macro recall (macro-Rec), macro precision (macro-Prec), and macro f1-score (macro-f1).

• 

  Figure 10. Receiver operating characteristic (ROC) curves obtained when classifying the full-length (10-min long) LDV signals with Support Vector Machines (SVM).

• 

  Figure 11. An example of the heart-rate (HR) computed from an SVM classifier on the full-length (10 min long) LDV signal and from ECG in one subject.

• 

  Figure 12. Boxplot of the absolute differences between the mean value of the heart rate (HR) computed with the LDV and the ECG in the 28 subjects.