Background: Cancer and cardiovascular disease are the leading causes of death worldwide. Cardiotoxicity, the intersection of these conditions, poses challenging health threats, and thus, its early prediction and prevention can drastically improve patient outcomes. However, accurate predictive methods remain elusive. This research aims to develop a first-in-class machine learning model to predict chemotherapy-induced cardiotoxicity using a novel biomarker, p16INK4a , and routine clinical data. Methods: A deidentified, institutional review board–approved cancer database was preprocessed using data imputation, normalization, and label encoding for model development. Machine learning models were developed with 56 clinical data points and predicted major cardiovascular adverse events such as systolic heart failure, pulmonary edema, myocardial infarction, and stroke. Using Python, regression and classification algorithms were fitted to the training data and evaluated by cross-validation. Results: After model optimization with hyperparameter tuning and refined feature selection, a multioutput logistic regression model was built and achieved clinically applicable accuracy in predicting cardiotoxicity, with a sensitivity/specificity of 92.2%/92.1% and positive/negative predictive values of 96.6%/81.6%. Composite MACEs were best predicted with a sensitivity/specificity of 86.49%/57.26%. Biomarker p16INK4a was among the most important predictors in these models. Conclusion: In this study, a novel biomarker was identified as a predictor of cardiotoxicity risk. To the best of our knowledge, we developed the first biomarker-based, personalized cardiotoxicity risk predictive machine learning model with the accuracy and readiness for clinical application, an important step toward optimized cardiotoxicity prevention in cancer patients. Further model training using a large dataset may provide an effective clinical decision-making tool for cardio-oncology care.
Liu et al. (Fri,) studied this question.