Time and event-specific deep learning for personalized risk assessment after cardiac perfusion imaging

Figures

• 

  Fig. 1. Deep learning enabled time-to-event outcome prediction after cardiac imaging - study overview. a A time-to-event deep learning model was trained (left) using data from the 5 sites of the REFINE SPECT registry (n = 20,401), then tested internally in a 10-fold cross-validation regimen (middle) and tested in 3 external sites (n = 13,988) (right); b The time-to-event model uses 5 SPECT polar maps and 15 clinical features as inputs (left) and predicts time-dependent probability of death (orange line), ACS (green line), and revascularization (blue line); c The performance of the model (left) is analyzed using cumulative dynamic area under the receiver-operating curves (cAUC). Red line represents the time-to-event model and blue line represents perfusion abnormality. The explanation of the prediction is visualized as a waterfall plot with blue arrows representing features that decrease the risk and red arrows representing the features that increase the risk (right); ACS acute coronary syndrome, AUC area under the receiver operating characteristics curve, TPD total perfusion deficit, PCI percutaneous coronary intervention, CI confidence intervals.

• 

  Fig. 2. Cumulative dynamic areas under the receiver operating characteristic curve (cAUC) in the internal testing set. The cAUCs are presented for the prediction of any adverse event (a) as well as specific events: all-cause death (b), acute coronary syndrome (c), and revascularization (d) in the internal testing dataset. Results were generated separately for each of the 10 testing folds of the cross-validation (N = 2041) and then averaged. The red, blue, and orange lines represent cAUC for the time-to-event model, perfusion abnormality, and clinical-only model, respectively. The red, blue, and orange dashed lines represent mean cAUC for the time-to-event model, perfusion abnormality, and clinical-only model, respectively. Shaded areas represent 95% confidence intervals (CI). The red arrows show the cAUC values of time-to-event deep learning averaged over the first six months after the scan (vertical dashed line).

• 

  Fig. 3. Cumulative dynamic areas under the receiver operating characteristic curve (cAUC) in the external testing set. The cAUCs are presented for the prediction of any adverse event (a) as well as specific events: all-cause death (b), acute coronary syndrome (c), and revascularization (d) in the external dataset (N = 13,988). The red, blue, and orange lines represent cAUC for the time-to-event model, perfusion abnormality, and clinical-only model, respectively. The red, blue, and orange dashed lines represent mean cAUC for the time-to-event model, perfusion abnormality, and clinical-only model, respectively. Shaded areas represent 95% confidence intervals (CI). The red arrows show the cAUC values of time-to-event deep learning averaged over the first six months after the scan (vertical dashed line).

• 

  Fig. 4. Patient-level prediction of time-dependent risk of major adverse cardiovascular events. Individual prediction of event probability in 4 different patients: (a) A 58-year-old male with no history of CAD, stress TPD of 2%, and diabetes; (b) A 76-year-old male with history of PCI and CABG, stress TPD of 3%, and family history of CAD and dyslipidemia; (c) A 63-year-old male with no history of CAD, stress TPD of 20%, and no other risk factors; (d) A 60-year-old male with no history of CAD, stress TPD of 1% and no risk factors. Curves represent predicted cumulative event probability of death (orange lines), acute coronary syndrome (green lines), and revascularization (blue lines) as a function of time. The vertical dashed lines mark the time of the true event. CABG – coronary artery bypass grafting, CAD coronary artery disease, PCI percutaneous coronary intervention, TPD total perfusion deficit (measure of perfusion abnormality).

• 

  Fig. 5. Patient-level risk explanation with simulated modification of risk factors. a Polar map visualization of perfusion, motion, and thickening of a 41-year-old female with no history of CAD, LVEF of 39%, and BMI of 31. Volume and degree of affected myocardium (0-100%) is visualized using color on the polar map, as shown on a color legend. b Waterfall plots present the influence of top features (including averaged contributions from polar maps) on prediction of average risk of death. The plot starts at the top at the mean population probability (expected value, dotted vertical line) and ends at the bottom with the final individual probability. Red arrows pointing to the right indicate that the feature increases the risk of death, while blue arrows pointing to the left indicate that the feature decreases the risk. Patient features that point to potential actionable interventions through medication or lifestyle modification are highlighted using bold font. c Simulated modification of risk factors through reduction of resting heart rate to 70/min reduces the average risk of death by 36%. The predicted probabilities of death, ACS and revascularization before the intervention is shown as…