Novel biomarker-based machine learning model in personalized cardiotoxicity prediction

Abstract Background Every year, 20 million cases of cancer are diagnosed globally. By 2026, it is estimated that there will be ~20 million cancer survivors in the US alone. Among cancer patients and survivors, cardiotoxicity is the leading cause of non-cancer death. Early prediction of cardiotoxicity is crucial for the timely initiation of cardio-protective strategies and optimized cancer treatment regimens. However, current methods of cardiotoxicity assessment are limited in their ability to predict early cardiotoxicity prior to treatment. Purpose This research aims to develop machine learning models for pre-treatment cardiotoxicity prediction, primarily utilizing peripheral blood p16 expression levels—a novel biomarker for cellular senescence and cardiotoxicity risk—along with routine clinical/demographic data. Methods A dataset of 185 de-identified individuals from the NORDICA clinical trial was pre-processed using data imputation, normalization, and label encoding for model development. Using Python, baseline machine learning algorithms were trained with comorbidities, demographics, risk factors, and lab data to predict major cardiovascular adverse events (MACEs), including cardiotoxicity, pulmonary edema, myocardial infarction, and stroke. Their validation performances were compared based on various evaluation metrics to select the best-performing algorithm. The highest-performing baseline model was then optimized with hyperparameter tuning and evaluated with 5-fold cross-validation for the average weighted F1 score, area under the receiver operating characteristic curve, and diagnostic accuracy metrics. Results The multi-output logistic regression model was developed and achieved clinically applicable accuracy in predicting cardiotoxicity by 56 clinical data points, with a sensitivity and specificity of 92.22% ± 4.22% and 92.11% ± 7.51% and positive and negative predictive values of 96.58% ± 3.15% and 81.62% ± 10.71%. Composite MACEs were best predicted by a class-output logistic regression model with a sensitivity of 86.49% ± 5.43%, specificity of 57.26% ± 1.61%, positive predictive value of 84.04% ± 11.68%, and negative predictive value of 65.57% ± 3.23%, a level of accuracy that approximates general clinical diagnostic standards. The probabilistic model performed exceptionally in its discriminatory ability between the positive and negative class (MACEs/no MACEs) with an AUC of 0.92 ± 0.03. A web-based user interface with the integration of the multi-output logistic regression model was subsequently developed to assist clinical decision-making. Conclusions This research pioneers the application of ML in pre-treatment cardiotoxicity prediction, introducing a robust biomarker and clinical data-based predictive model with clinical applicability. This study additionally demonstrates that p16 is an essential biomarker for cardiotoxicity. Further model training in a larger database may allow the direct use of this tool in cardiotoxicity prevention.