Model-Driven Analysis of ECG Using Reinforcement Learning

Figures

• 

  Figure 1. Generation of the ECG PQRST complex. The left–right image shows the positioning of ECG electrodes. Seven smaller representations of the heart illustrate the direction and polarity of electrical dipoles generated at each heartbeat phase. The bottom of the figure displays the resulting PQRST waveform. Red (blue) arrows indicate the direction of depolarization (repolarization) waves.

• 

  Figure 2. Signal quality as a function of the SNR¯, as defined in (3). The left panels show 10 s of the original signals, whereas the right panels show all the normalized beats (in light grey) and the mean beat (red). From top to bottom, we display examples of recordings with an SNR¯ closest to −15, −10, −5, 0, 5, 10, and 15 dB.

• 

  Figure 3. This plot overlays the mean PQRST for each valid segment (light grey) and their median (cyan). We also overlaid the sigma-lognormal prototype (red) and the upper (dashed red) and lower (dashed blue) envelop of the PQRST time series that can be generated by the model when using parameters within the ΣΛ− and ΣΛ+ bounds specified in Table 1. We also show the space covered by each component in different colors.

• 

  Figure 4. Schematic representation of the interaction between the agent and the environment in the context of reinforcement learning.

• 

  Figure 5. Network architecture used for deep reinforcement learning.

• 

  Figure 6. Fitting SNR. (a) Distribution of fitting SNR for individual beats. (b) Twelve randomly selected beats (solid blue) with their fitting (dashed orange) and the corresponding SNR.

• 

  Figure 7. Correlation between the age and the modeling parameters. The light blue region denotes the 95% confidence interval obtained with bootstrapping. Overlaid on each panel is the Kendall rank correlation coefficient (τ) and its corresponding p-value corrected for 24 independent tests using Bonferroni’s correction (i.e., reported p-values have been multiplied by 24), both in red if significant (i.e., p<0.01). Parameters D are expressed in μV ∗ s (i.e., multiplied 1×106), for convenience.

• 

  Figure 8. Comparison of PQRST complexes predicted from Sigma–Lognormal modeling (top panel), mean (middle panel), and median (bottom panel) PQRST across participants, for each time point with at least 8 participants. Both columns show the same data; the right columns only zoom further on the P peak to better appreciate the modulation of its amplitude by age.