<p>Immune Checkpoint Inhibitor (ICI)-related cardiotoxicity has a high mortality rate, making early prediction crucial for improving patient prognosis. However, early prediction models are currently lacking in clinical practice. This study aims to develop an early prediction model for ICI-related cardiotoxicity using the eXtreme Gradient Boosting (XGBoost) algorithm. Retrospective analysis was conducted on patients who received ICI therapy between January 2020 and December 2023. The population was categorized into a cardiotoxicity group and a non-cardiotoxicity group based on the presence of cardiac biomarkers and electrocardiogram abnormalities that could not be attributed to other diseases within 30 days after initiation ICI therapy. The dataset was split into training (70%) and testing (30%) sets. Logistic Regression (LR), Random Forest (RF), and XGBoost models were constructed in Python, with variables selected based on each model’s characteristics. The models were compared based on predictive performance, which was measured by area under the curve (AUC) and decision curve analysis (DCA). The best model was explained using SHapley Additive exPlanation (SHAP). A total of 419 patients were included. The XGBoost model demonstrated the highest predictive performance with an AUC of 0.83, outperforming LR (AUC: 0.80) and RF (AUC: 0.74) models. DCA confirmed the XGBoost model’s superior net benefit. Among the selected predictors, cardiac troponin T (cTnT) emerged as the most important variable, demonstrating the highest feature importance. The XGBoost model proposed could assist clinicians in personalized risk stratification for patients on ICI therapy, facilitating precise monitoring of cardiotoxicity and tailored treatment strategies.</p> Graphical Abstract <p>Overview of the study: A. Research Purpose: A machine learning model was developed utilizing readily available baseline data to predict the occurrence of cardiotoxicity in patients within 30 days of initiating immune checkpoint inhibitor (ICI) therapy. B. Model Establishment and Evaluation Process: The data were divided into a training set and a testing set. The training set was employed to screen the feature variables for each model and subsequently train the model. The testing set was then utilized to further assess the model’s generalization ability. Model performance was compared using the Receiver Operating Characteristic (ROC) curve and Decision Curve Analysis (DCA), with XGBoost demonstrating the highest performance among the three models. C. Best Model Result Interpretation: The selected variables were input into the XGBoost model, and SHAP (SHapley Additive exPlanations) was used to elucidate the model’s decision-making process. In the SHAP analysis, red features indicate that higher baseline values correspond to an increased likelihood of cardiotoxicity (e.g., cTnT, SBP, RV5, hs-CRP, and MYO), while blue features indicate the opposite. Notably, cTnT exhibited the most significant impact on the XGBoost model.</p> <p></p>