Establishment and validation of an electrocardiogram vector-based machine learning model for the conversion of prone position electrocardiograms into standard electrocardiograms

Figures

• 

  Figure 1. The importance of features for two machine learning (ML) models in different segments (QRS complex, ST-T segment, P-R segment) of leads V1–V5. (A–E) Machine learning models for QRS complex of leads V1–V5. (F–J) Machine learning models for ST-T segment of leads V1–V5. (K–O) Machine learning models for P-R segment of leads V1–V5. The y-axis represents the variables included in the models, with the variables contained in the vertical line indicating those ultimately selected in the random forest (RF) model. The x-axis represents the importance of each variable, where bars on the right indicate the variable importance in the extreme gradient boosting (XGBoost) model, and bars on the left represent the variable importance in the random forest model.

• 

  Figure 2. Comparison of evaluation indices for three approaches. The chart illustrates a comparison of evaluation indices for three approaches, with the x-axis representing root mean square error (RMSE), the y-axis representing determination of coefficient [R-squared (R2)], and marker size indicating mean absolute error (MAE). Models positioned closer to the upper-left corner with smaller markers exhibit better performance. Approaches 1 and 2 calculate model evaluation indices for healthy volunteers and cardiovascular disease, respectively. While Approach 3 calculates evaluation indices for the testing group.

• 

  Figure 3. Comparison of amplitudes and morphology in electrocardiograms (ECGs) of external validation dataset after transformation using each approach and comparison with the original electrocardiograms. (A) Comparison of average amplitudes of P waves and T waves in converted electrocardiograms with the original electrocardiograms. (B) Comparison of average amplitudes of R waves and S waves in converted electrocardiograms with the original electrocardiograms. In (A and B), *, indicates a significant difference in amplitude compared to the original electrocardiogram (P < 0.05), # indicates a significant difference in amplitude compared to the electrocardiogram converted using the Approach 1 (P < 0.05), † indicates a significant difference in amplitude compared to the electrocardiogram converted using the Approach 2 (P < 0.05), ‡ indicates a significant difference in amplitude compared to the electrocardiogram converted using the Approach 3 (the RF method) (P < 0.05), and § indicates a significant difference in amplitude compared to the electrocardiogram converted using the Approach 3 (XGBoost method) (P < 0.05). (C) Comparison of morphology similarity in each segment of converted…

• 

  Figure 4. Comparison of electrocardiograms (ECGs) converted from different types of prone electrocardiograms using hybrid models in the external validation set with the original standard electrocardiograms. (A) Normal electrocardiogram; (B) ST-segment depression; (C) Bundle branch block (complete right bundle branch block); (D) old anterior wall myocardial infarction; (E) ST-segment elevation.

• 

  Figure 5. The diagnoses comparison between original electrocardiograms (ECGs) and converted standard electrocardiograms diagnosed by experts in the external validation dataset (A) and the predictive performance of the converted standard electrocardiograms for each type of electrocardiograms (B). (A) The left y-axis represents the diagnoses of each original electrocardiogram in the external validation dataset, while the right y-axis shows the diagnoses provided by experts for the electrocardiograms converted through the model. (B) Receiver operating characteristic (ROC) curves for the diagnoses of the conversed electrocardiograms based on expert assessments are displayed. The bar graph illustrates the sensitivity, specificity, area under the receiver operating characteristic curve (AUC), and F1 score of the models for various disease. MI, myocardial infarction.