Rethinking multiscale cardiac electrophysiology with machine learning and predictive modelling

Abstract

We review some of the latest approaches to analysing cardiac electrophysiology data using machine learning and predictive modelling. Cardiac arrhythmias, particularly atrial fibrillation, are a major global healthcare challenge. Treatment is often through catheter ablation, which involves the targeted localised destruction of regions of the myocardium responsible for initiating or perpetuating the arrhythmia. Ablation targets are either anatomically defined, or identified based on their functional properties as determined through the analysis of contact intracardiac electrograms acquired with increasing spatial density by modern electroanatomic mapping systems. While numerous quantitative approaches have been investigated over the past decades for identifying these critical curative sites, few have provided a reliable and reproducible advance in success rates. Machine learning techniques, including recent deep-learning approaches, offer a potential route to gaining new insight from this wealth of highly complex spatio-temporal information that existing methods struggle to analyse. Coupled with predictive modelling, these techniques offer exciting opportunities to advance the field and produce more accurate diagnoses and robust personalised treatment. We outline some of these methods and illustrate their use in making predictions from the contact electrogram and augmenting predictive modelling tools, both by more rapidly predicting future states of the system and by inferring the parameters of these models from experimental observations.

Keywords: Cardiac arrhythmia; Cardiac electrophysiology; Deep learning; Electrogram; Machine learning; Predictive modelling.

Fig. 1

Examples of 20 ms segments of electrogram recordings from the Control and CBX groups, after removal of the stimulus artefact.

Fig. 2

3D scatter plot of the most relevant features, normalised in the interval [0,1], as determined by SFS. No single feature clearly discriminates between the control and carbenoxolone classes.

Fig. 3

Confusion matrix comparing the predictability of classes using the training dataset. Diagonal cells (green) show the percentage of electrograms that were correctly classified.

Fig. 4

Schematic of the convolutional neural network.

Fig. 5

Confusion matrix from the binary classification of electrograms before and after administration of carbenoxolone to monolayers of cultured myocytes.

Fig. 6

The architecture of the network used to predict the future behaviour of the 2D diffusion system.

Fig. 7

Accuracy of next steps prediction versus number of input ($K_b$) and back-propagated ($K_t$) frames. Mean Squared Error (MSE), averaged over the test dataset, first, and the 5 cross-validation folds, then, against $K_t$ and for $K_b=\mathrm{2,3}$. Error bars extend to the minimum and maximum MSE among the 5 folds. Dashed lines represent the last input level.

Fig. 8

Comparison between the full MSE distribution across the test dataset for the predictions of a single trained network, compared with that given by using the last input frame as the prediction. The asterisks represent statistical significance with p-value $<{10}^{-4}$ (Wilcoxon signed rank test).

Fig. 9

The logarithm of the Normalised Mean Squared Error (NMSE) against the prediction time for networks trained with various $K_t$ (and $K_b=3$). The black dots represent the $K_t$ corresponding to each line.

Fig. 10

Accuracy of parameter inference from the internal representation of the prediction network with $K_t=4$ and $K_b=3$ for (a) boundary angle θ, (b) boundary position β and (c–d) scar parameters $d_0$, $d_1$, $d_{0,scar}$ and $d_{1,scar}$. The predicted parameter values are plotted against the target values for a subset of the test dataset.

Fig. 11

(a) Kernel density estimates for steady-state activation parameters showing sequential constraining of distributions. (b) Data used to fit channel (black crosses) are plotted for each patch clamp experiment. Simulation results are also plotted for original parameter settings (green triangles), 100 samples from prior distribution (blue circles), and 100 samples from posterior distribution (orange squares). Shaded area indicates 95% confidence intervals around the median line.