Machine learning integration of echocardiographic and clinical data to improve prediction of survival following myocardial infarction

Figures

• 

  Figure 1. Patient flow data, with inclusions and exclusions, as well as splits for prediction algorithms.

• 

  Figure 2. Receiver operator curves (ROC) of all three models. Delong's P-value statistical test is used to compare AUC. NN, Neural Network model; XGB, XGBoost model.

• 

  Figure 3. Kaplan–Meier curves of predicted probability score of death by XGBoost model stratified into high- and low-risk groups. Youden's J statistic: 0.576, optimal cut-off threshold: 0.306.

• 

  Figure 4. XGBoost ShAP Beeswarm plot. This plot visually represents the influence of each variable on the XGBoost model's mortality prediction. Features are ranked by importance, with the most influential at the top. Each dot represents a patient, coloured by the feature's value (red = high, blue = low). Horizontal displacement from the centre indicates the feature's impact: rightward shift increases mortality risk, leftward shift decreases risk.

• 

  Figure 5. XGBoost ShAP Waterfall plot. Visual representation of two selected patients and how the model used the attributes to calculate their final score. Beginning from the mean prediction at the bottom (ML score of 0.442), the waterfall plots display SHAP (SHapley Additive exPlanations) values in increasing order of magnitude, with red for increasing risk and blue for decreasing the risk of death. The positive or negative SHAP value of each feature is progressively added to reach the individual prediction. Case 1 shows the predictions for a 70-year-old patient post MI with an LVEF of 28.9%, a mitral E point velocity of 118 and an eGFR of 46, with a final ML score of 1.206 (high risk of death). In such a patient, this high score could inform subsequent changes in management. Case 2 shows a 73-year-old patient post MI who had an LA area of 14.6, eGFR of 67, an LA volume index of 19.5, and an LV e’ septal velocity of 6.2, resulting in a final ML score of −0.71 (low risk of death). In this example, based on the ML score, a different approach to management could be taken.

• 

  Figure 6. Receiver-operating curves for nested XGBoost models with/without comprehensive echocardiographic data. Delong's P-value statistical test is used to compare AUC.

• 

  Figure 7. Contribution of key echocardiographic parameters to ML models for predicting survival: schematic showing key echocardiographic predictors contributing significantly to ML prediction of survival, alongside clinical variables. LVOT, left ventricular outflow tract; VTI, velocity time integral; LVEF, left ventricular ejection fraction; LV, left ventricular; MI, myocardial infarction.