Fig. 4: SSCAR interpretation.
From: Arrhythmic sudden death survival prediction using deep learning analysis of scarring in the heart
The features learned by SSCAR are interpreted by performing a gradient-based sensitivity analysis of the location parameter (the most probable TSCDA) to changes in the neural network input or features. The gradient value quantifies this sensitivity. The magnitude of the gradient measures the strength of the sensitivity of the predicted TSCDA to inputs or intermediary features. The sign of the gradient shows the direction of the effect. That is, for a small increase in the value of inputs or features, a positive gradient (blue) indicates a higher predicted TSCDA, whereas a negative gradient (red) indicates a decrease in the predicted TSCDA. a, Shown is the CMR sub-network feature interpretation for an example patient who did not experience SCDA (No SCDA, top) and for a patient who did (SCDA, bottom). For each patient, a subset of 3 of the 12 contrast-enhanced short-axis CMR images (corresponding to three locations in the heart, base to apex, top to bottom, left column) used as inputs by SSCAR are overlaid with blood pool and myocardium segmentation (middle column, orange and green, respectively). A heat map of extracted features scaled by the value of the gradient shows contribution of the local pixel intensity to the predicted location parameter for the last convolutional layer (right column, blue and red heat maps). Of note, although the patient with SCDA shows high gradients in areas with contrast enhancement, the patient on the left shows that enhancement can also lead to positive gradients, suggesting that the network does not simply create a mask of the enhanced regions to make predictions but learns a nuanced relationship between scar and propensity for SCDA. b, Covariate sub-network interpretation based on an average of all patients (mid-blue and mid-red bars), patients with SCDA (dark blue and dark red bars) and patients with no SCDA (light blue and light red bars). Top four highest (blue bars) and bottom four lowest (red bars) average gradients of the neural network output (that is, the predicted location parameter) with respect to the clinical covariate inputs are shown. The error bars represent approximate 95% CIs. LVEF CMR, left ventricular ejection fraction computed from CMR; betablock, use of β-blocker medication; ECG hr, heart rate from ECG; digoxin, use of digoxin medication; infarct %, infarct size as % of total volume; ECG QRS, QRS complex duration from ECG; LV mass ED, left ventricular mass in end-diastole.
