Machine learning-based prediction model for myocardial ischemia under high altitude exposure: a cohort study

Figures

• 

  Figure 1. Workflow for data management and myocardial ischemia prediction model development. Data came from 4000 adults who completed physical exams at the 920th Hospital of Joint Logistics Support Force between January and June 2022 and prepared to enter the plateau within 6 months. After exclusion, there were 2855 people remaining. Participants' data (n = 2855) were randomly assigned to a prediction model training dataset (n = 2141) and a test dataset (n = 714) in a 3:1 ratio following preprocessing. fivefold cross-validation was used for training and selecting the prediction model, and five classification algorithms were evaluated. Feature selection was conducted using the RFE (Recursive Feature Elimination) algorithm. The final prediction model was validated using a test dataset. LR logistic regression, RF random forest, XGBoost eXtreme gradient boosting, KNN K-nearest neighbor, SVM support vector machines.

• 

  Figure 2. Results from the model training and algorithm selection process. (A) The bar chart shows the average AUC score for each candidate algorithm in the training set. (B) The bar chart shows the AUC score of the best model for each candidate algorithm in the test set. (C) The best ROC curve for each candidate algorithm in the training set. (D) ROC plot of the best model for each candidate algorithm in the test set. Details of these analysis methods are provided in the appendix (Additional file 2: Table S2). AUC area under the receiver operating characteristic curve, ROC Receiver operating characteristic (ROC) curve.

• 

  Figure 3. Variable importance scores of RF, XGBoost and LR. (A)–(C) Variable importance scores of RF, XGBoost and LR, respectively. Details of these analysis methods are provided in the appendix (Additional file 3: Table S3). (D) The Venn diagram of the top 50% ranked variables in the RF, XGBoost and LR algorithm models, with 9 intersecting variables including Eos.Per, G, Ca, PLT, GLU, AST, TP, MCHC, and Crea.

• 

  Figure 4. AUC versus the number of variables used in RFE and performance comparison of 27F model and 5F model under RF algorithm. (A) The performance of RFE with respect to different variables using the area under the ROC curves. The graph illustrates that the model achieved comparable performance with the highest AUC by utilizing 5 selected features. (B) The ROC curves of the 27F and 5F models based on the training dataset. On (C), we have shown the ROC curves for the 27F and 5F models for the testing dataset. The calibration curves for the 5F and 27F models for the training dataset (D) and test dataset (E) are presented.