Left Ventricular Thrombus in Ischemic Heart Failure: Machine-learning–based Prediction of Six-month Persistence and One-year Outcomes

Figures

• 

  Fig. 1. Panel A–C. Summary of prediction displays. (A) Predicted 1-year MACE vs. thrombus size (marginal). (B) Predicted 1-year MACE vs. CHA₂DS₂-VA score (marginal). (C) Thrombus size vs. predicted 1-year MACE stratified by CHA₂DS₂-VA level (conditional)

• 

  Fig. 2. Global SHAP feature importance—Top-20 baseline predictors (model-agnostic display). Overall ranking of the top 20 baseline predictors by mean(|SHAP value|). The list again prioritizes left atrial diameter, pulmonary artery pressure, platelet count, and LVEDD, with metabolic/renal markers (e.g., triglycerides, creatinine, sodium) contributing smaller but consistent signals. Results are averaged on the evaluation split; details of imputation, scaling, encoding, and resampling are provided in Methods

• 

  Fig. 3. CatBoost—Global SHAP feature importance for 6-month LV thrombus non-regression. Global importance ranked by mean(|SHAP value|) on the hold-out data; higher bars indicate greater average contribution to the model’s predictions of non-regression (Group 2). Top predictors were left atrial diameter, platelet count, pulmonary artery pressure, and LVEDD, whereas CHA₂DS₂-VA score and thrombus size (area method) ranked lower—highlighting complementary hemodynamic/remodeling and hematologic signals beyond thrombus burden and clinical risk score. Pre-processing and cross-validation are described in Methods

• 

  Fig. 4. CatBoost—SHAP summary (beeswarm) plot for 6-month LV thrombus non-regression. Each dot represents one patient; the x-axis shows the SHAP value (impact on the log-odds of non-regression), with positive values shifting the prediction toward non-regression and negative values toward regression. Dot color encodes the feature value (red = high, blue = low). Higher left atrial diameter, pulmonary artery pressure, LVEDD, and platelet count cluster to the right (risk-increasing), whereas higher ejection fraction tends to shift predictions left (protective). This pattern supports a multidimensional risk phenotype (stasis/remodeling + hypercoagulability)