Prediction models for immune checkpoint inhibitor-related cardiovascular toxicity: a systematic review.

Background: Immune checkpoint inhibitor (ICI)-related cardiovascular toxicity is associated with high mortality, and prediction models may facilitate its prevention and management. However, the predictive performance, quality, risk of bias, and applicability of existing models remain unclear. Objective: To systematically review prediction models for ICI-related cardiovascular toxicity. Methods: Nine databases were searched for studies on prediction models for ICI-related cardiovascular toxicity from inception to July 12, 2026. Data were extracted in accordance with the Checklist for Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies. The included models were assessed using the Quality, Risk of Bias, and Applicability Assessment Tool for Prediction Models Using Regression or Artificial Intelligence Methods. This systematic review was registered on PROSPERO (CRD420261296378). Results: A total of 34, 541 articles were identified, and 23 articles met the inclusion criteria. These 23 articles described 82 prognostic models (71 for predicting the occurrence risk of ICI-related cardiovascular toxicity and 11 for predicting the prognosis of this toxicity) and two diagnostic models. Modeling methods comprised logistic or Cox regression and machine learning algorithms. Common predictors included demographic characteristics, medical history, treatment-related characteristics, and laboratory and imaging parameters. The area under the receiver operating characteristic curves or C statistic ranged from 0.699 to 0.967 for apparent performance, 0.595 to 0.942 for internal performance, and 0.549 to 0.949 for external performance. Calibration and clinical utility were assessed in 20 and 15 models using calibration curves/Hosmer-Lemeshow test and decision curve/clinical impact curve analysis, respectively, in apparent, internal, or external performance validation. Models were presented using techniques such as nomograms, scoring systems, and regression equations. All models were graded as low quality, high risk of bias, and only seven models were graded as high applicability. Conclusions: Existing prediction models have concerns regarding quality, risk of bias, and applicability, limiting their clinical practice. Currently, baseline and dynamic risk screening using relevant risk biomarkers is suggested to better manage ICI-related cardiovascular toxicity. Future research is warranted to validate and improve existing models or develop new prediction models with rigorous methods. Systematic review registration: https://www.crd.york.ac.uk/PROSPERO/, identifier CRD420261296378.