Improving arrhythmic risk prediction using cardiac magnetic resonance within deep learning in ischemic heart disease

Figures

• 

  Fig. 1. Feature inclusion across survival models and dataset configurations using clinical variables. The heatmap shows the selection of individual features for CoxPH, RSF, and DeepSurv models under three train--test configurations. Rows correspond to candidate predictors, and columns represent model--configuration combinations. Dark cells indicate selected features, whereas light cells indicate features that were not selected.

• 

  Fig. 2. Feature inclusion across survival models and dataset configurations using clinical and imaging variables. The heatmap shows the selection of individual features for CoxPH, RSF, and DeepSurv models under three train--test configurations. Rows correspond to candidate predictors, and columns represent model--configuration combinations. Dark cells indicate selected features, whereas light cells indicate features that were not selected.

• 

  Fig. 3. Calibration of survival predictions using clinical variables. Calibration curves are shown for CoxPH, RSF, and DeepSurv across three dataset configurations (Conf-1, top left; Conf-2, top right; and Conf-3, bottom left). Solid, dashed, and dotted lines represent 2-, 5-, and 8-year predictions, respectively. The diagonal dashed line denotes perfect calibration.

• 

  Fig. 4. Risk stratification using model-predicted survival probabilities using clinical variables. Kaplan--Meier survival curves stratified patients into low-, intermediate-, and high-risk groups according to predicted risk scores generated by the Cox proportional hazards (CoxPH), random survival forest (RSF), and DeepSurv models. Results are shown for the three dataset configurations (Conf-1, Conf-2, and Conf-3). Shaded regions represent 95% confidence intervals.

• 

  Fig. 5. Calibration of survival predictions using clinical and imaging variables. Calibration curves are shown for CoxPH, RSF, and DeepSurv across three dataset configurations (Conf-1, top left; Conf-2, top right; and Conf-3, bottom left). Solid, dashed, and dotted lines represent 2-, 5-, and 8-year predictions, respectively. The diagonal dashed line denotes perfect calibration.

• 

  Fig. 6. Risk stratification using model-predicted survival probabilities using clinical and imaging variables. Kaplan--Meier survival curves stratified patients into low-, intermediate-, and high-risk groups according to predicted risk scores generated by the Cox proportional hazards (CoxPH), random survival forest (RSF), and DeepSurv models. Results are shown for the three dataset configurations (Conf-1, Conf-2, and Conf-3). Shaded regions represent 95% confidence intervals.

• 

  Fig. 7. Outcome characteristics of the study populations. (left) Histogram of time to major arrhythmic events in Dataset 1 (solid line) and Dataset 2 (dashed line), shown using 200-day intervals. (right) Kaplan--Meier curves illustrating arrhythmia-free survival in the two datasets. Solid and dashed lines correspond to Dataset 1 and Dataset 2, respectively.

• 

  Fig. 8. LGE-CMR image processing and feature extraction pipeline. Representative short-axis late gadolinium enhancement cardiovascular magnetic resonance images are segmented using the full-width at half-maximum (FWHM) method to delineate infarct scar and peri-infarct zone (PIZ) regions. Morphological and texture-based features are subsequently extracted for downstream analysis.

• 

  Fig. 9. Survival analysis workflow for arrhythmic risk prediction. The methodological pipeline includes data preprocessing, feature extraction, feature selection, model training, and performance evaluation using multiple train--test configurations. Survival prediction was performed using Cox proportional hazards regression (CoxPH), random survival forest (RSF), and DeepSurv models.