Prediction of Myocardial Infarction From Patient Features With Machine Learning

Figures

• 

  Figure 1. A typical Delayed Enhancement MRI (DE-MRI) slice involving the myocardial infarction (MI) and the persistent microvascular obstruction (PMO). On the left the cropped short-axis DE-MRI covering the whole left ventricle, on the right the corresponding masks of normal myocardium (green), MI (blue, PMO exclusive) and PMO (yellow). The PMO is semi-wrapped by the MI and contacts the cavity.

• 

  Figure 2. Workflow of the proposal. This figure presents the inference stage for the quantification task and the classification task of the automatic MI prediction. On the left part, selected patient features are first preprocessed to 17 pieces of numerical or Boolean features. For the quantification task, the features are incorporated through a regression model so that the obtained value is the Percentage of Infarcted Myocardium (PIM) ranging from 0 to 1. In the case of the classification task, the prediction can be obtained by either a regression model followed by thresholding or a classification model. During the training stage, the regression model is supervised by the ground truth PIM, and the classification model is supervised by the ground truth state of the myocardium. Both ground truths are defined from the DE-MRI and manual annotations.

• 

  Figure 3. Receiver operating characteristic curves of classification results. The classification of different tissues was realized with Random Forest (RF) algorithms with different optimization functions.

• 

  Figure 4. Impact of the training set volume on the mean and the SD of quantification error. The dataset of 150 cases was randomly split into multiple folds to have different amounts of training data. The training set volume ranged from 10 to 149 for each cross-validation.

• 

  Figure 5. Prediction error on cases of different severity according to the infarction. Normal cases were not considered in this figure. SD, Standard Deviation; mean diff, mean difference.