Machine Learning Electrocardiogram for Mobile Cardiac Pattern Extraction

Figures

• 

  Figure 1. System diagram of novel machine learning framework, which identifies the heartbeats with the support vector machine (SVM) in stage I, localizes the QRS complexes with the multiview dynamic time warping (MV-DTW) in the first part of stage II, purifies the QRS complexes in the second part of stage II, and then transforms the mobile-ECG-based QRS durations to commonly used standard chest-ECG-based ones in stage III. The blocks above the dashed line are for training, and those under the line are for testing, so the HB ID model, HB template learning, and distortion threshold learning are only included in the training phase. Notes: Orange blocks: supervised learning steps; blue blocks: unsupervised learning steps; green blocks: pattern recognition steps; HB: heartbeat; ID: identification; SVM: support vector machine; DTW: dynamic time warping.

• 

  Figure 2. The mobile ECG (c) owns distinguishable morphologies, but is much weaker and more sensitive to motion artifacts, compared with the traditional chest-ECG (a). P, R, R, S, and T, are characteristic points of an ECG heartbeat. The on and off positions of the QRS complex are also labeled in the chest-ECG zoomed in (b), and the corresponding mobile ECG zoomed in is also given (d) to demonstrate the mobile ECG is of a very low bio-potential and sensitive to noise.

• 

  Figure 3. An example of the SVM-based heartbeat identification results (a), corresponding to the second minute data of one trial in the testing session of a subject (subject 6 is selected as an example), and showing the necessity to deal with many highly distorted heartbeats and some faking heartbeats. Two segments are further visualized with more details in (b,c). The red dots are detected heartbeat R peak locations.

• 

  Figure 4. An example of a representative heartbeat learned by k-medoid clustering, corresponding to the training session of subject 6, where the medoid in cluster 3 is selected as a high-quality template since it represents the highest number of raw heartbeats (51%) among three medoids. Dark to blue lines: different dimensions of the three-dimensional raw heartbeats; bold lines: different dimensions of the medoids; all dimensions in all raw heartbeats are scaled to be between 0 and 1. Notes: Amp—amplitude; Deri—derivative.

• 

  Figure 5. An example of the multiview-DTW-based nonlinear stream matching, corresponding to a raw heartbeat in the testing session of subject 6, and showing that the QRS offset/onset in the three-dimensional test stream is robustly located using a backward search method referring to the locations in the template stream. To determine the off and on boundaries of the QRS complex of the mobile ECG in the bottom, the corresponding preknown off and on locations in the chest ECG template on the left side are firstly mapped to the optimal warping path in the middle graph and then mapped to the mobile ECG in the bottom. Notes: the yellow/blue colors correspond to the highest/lowest path distance values, respectively. Amp—amplitude; Deri—derivative.

• 

  Figure 6. An example of the histogram triangle-based distortion threshold learning (a), corresponding to the training session of subject 6, and showing that the transition point (greed dot on the left side) is learned to determine an appropriate distortion threshold (0.25 with normalization). The right plot (b) is based on the distance from each point on the right side of the red envelop to the blue line in the left plot, so each distance is actually corresponding to a perpendicular line to the blue line. The green dot in the right plot is maximum, meaning that there is a sharp transition in the red envelop in the left plot and this transition is a good threshold.

• 

  Figure 7. An example of the SQI-based raw heartbeat purification (a), corresponding to the testing session of subject 6, and indicating that segments highly distorted by motion artifacts have been successfully filtered out and other segments with a high quality are reserved. Red/green curves in (b): the smoothed distortion sequence and the distortion threshold, respectively. Two segments are further visualized to demonstrate the high-noise condition (d) and relative the normal-noise condition (e).

• 

  Figure 8. Correlation (Corr.) and Bland-Altman (BA) plots between predicted chest-QRS durations (QRSpred) and real chest-QRS durations (QRSchest), without (a,b) calibration (Cali.) model learning. Results is based on the testing sessions of all subjects; MDTW: MV-DTW. (a1) and (b1) give the correlation plots, and (a2,b2) give the BA plots.