Non-hyperaemic assessment of coronary ischaemia: application of machine learning techniques

Figures

• 

  Figure 1. Stepwise representation of the supervised machine learning algorithm. Panels labelled A–F (Steps 1–6) display the ability of including/excluding different prediction intervals of the cardiac cycle. Data points to the left of the dashed black lines in Row 1 belong to whole-cycle P(aortic) data, with data points to the right belonging to the 18–64% prediction interval of P(distal). Optimized mean classification accuracy, sensitivity, specificity, positive predictive value, and negative predictive value are calculated using the confusion matrix in the bottom right panel. ML, machine learning; NPV, negative predictive value; PPV, positive predictive value.

• 

  Figure 2. Principal component analysis biplot, ‘EigenCycles’ (Eigenvector) plot and loading values heatmap. The first four principal components in descending order of largest mean differences between positive and negative ischaemia principal component scores [P(aortic and distal) data using inter-vessel standardization and whole-cycle P(aortic) (0–100%) with the diastolic portion (18–64%) of P(distal) waveforms]. The EigenCycles plot and loading values heatmap shown in Column 2 are alternate ways of presenting the Biplot information. Data points to the left of the dashed black lines in Column 2 belong to whole-cycle P(aortic) data, with data points to the right belonging to the 18–64% prediction interval of P(distal). PCA, principal component analysis.

• 

  Figure 3. Unsupervised learning two-dimensional representations of multi-dimensional pressure waveforms produced using t-distributed stochastic neighbour embedding and Gaussian mixture modelling clustering. Using conditions, which yielded a supervised learning classification accuracy of 86.9% in Table 3 [ensemble average, inter-vessel, P(aortic and distal)]. Ischaemic labels are delineated by colour. The Gaussian mixture modelling clusters shown in these figures, which model data as a collection of Gaussian blobs in a probabilistic model describing the distribution of the data, were not able to find distinct clusters or separate ischaemic from non-ischaemic data. GMM, Gaussian mixture modelling; t-SNE, t-distributed stochastic neighbour embedding.

• 

  Figure 4. Unsupervised learning two-dimensional representations of multi-dimensional pressure waveforms produced using t-distributed stochastic neighbour embedding and Gaussian mixture modelling clustering. Comparison to Figure 3 using instead the individual cycle method and P(aortic–distal) over the entire cycle (0–100% prediction interval). Ischaemic labels are delineated by colour. The Gaussian mixture modelling clusters shown in these figures, which model data as a collection of Gaussian blobs in a probabilistic model describing the distribution of the data, were not able to find distinct clusters or separate ischaemic from non-ischaemic data. GMM, Gaussian Mixture Modelling; t-SNE, t-distributed stochastic neighbour embedding.

• 

  Figure 5. Range of fractional flow reserve values for the 77 coronary lesions. FFR, fractional flow reserve.