Machine learning cardiac-MRI features predict mortality in newly diagnosed pulmonary arterial hypertension

Figures

• 

  Figure 1. Model pipeline flow chart.

• 

  Figure 2. Study participants flow chart.

• 

  Figure 3. Kaplan–Meier curve. The Kaplan–Meier curve shows the survival of high- and low-risk patients based on the combined short-axis and four-chamber model predictions. The risk threshold was determined based on the median value of the MPCA predictions. The Kaplan–Meier analysis shows a significant difference in survival between the high and low risk of mortality patient groups (log-rank P < 0.001).

• 

  Figure 4. Receiver-operating characteristic curves for 1-year mortality prediction. The prognostic accuracy of the different machine learning models were compared (A) combined model, (B) short-axis model and (C) four-chamber model. The highest area under the curve was achieved with the short-axis model (AUC = 0.73).

• 

  Figure 5. Time-resolved prognostic cardiac features. Features of poor prognosis and also protective features were examined throughout the cardiac cycle on the short-axis and four-chamber views. The most significant cardiac features were the end-systolic and early diastolic septum on both the short-axis and four-chamber views. The RV during systole and LV during diastole also were predictive of 1-year mortality. In contrast, the most important features of survival were the end-diastolic septum on short-axis and four-chamber views and the LV at systole.

• 

  Figure 6. Visualization of prognostic mortality and survival features learnt from the training data. The features were overlaid on three example short-axis/4-chamber images from three different patients with PAH to interpret the corresponding anatomical regions. Left: septum and LV features of high risk of mortality at end-systole. Middle image: Features of survival visualized at the septum at end-diastole. Right: four-chamber view showing high-risk features in the septum and LV in diastole