Comparative analysis of machine learning models for predicting anthracycline-induced cardiotoxicity in patients with hematologic malignancies

Objective: a comparative study of various machine learning algorithms (AdaBoost, k-nearest neighbors, linear discriminant analysis, logistic regression, neural networks, random forest, stochastic gradient descent, support vector machines, XGBoost) for predicting anthracycline-induced cardiotoxicity (AIC) in patients with haematological cancer using clinical and instrumental predictors. Material and Methods. A prospective study included 155 haematological cancer patients receiving anthracycline-containing therapy. The age of the patients ranged from 18 to 74 years. Clinical data, biomarker levels (NT-proBNP, troponin I), and echocardiographic parameters of diastolic function (E’, E/E’, LAVI) were analyzed. Data underwent preprocessing (standardization, one-hot encoding), and class imbalance was addressed using SMOTETomek. Models were trained and evaluated via 5-fold stratified cross-validation using F1-score, AUC-ROC, precision, and recall metrics. Results. Statistically significant predictors of AIC included NT-proBNP (p<0.001), troponin I (p=0.004), and echocardiographic parameters E’ (p<0.001) and LAVI (p<0.001). Incorporating age and E/E’ ratio further enhanced the model predictive value. Logistic regression demonstrated optimal performance (F1 0.943 ± 0.070, AUC-ROC 0.963 ± 0.051) with perfect precision (1.00 ± 0.00) and high recall (0.90 ± 0.12). Linear discriminant analysis yielded comparable results (F1 0.921 ± 0.066, AUC-ROC 0.963 ± 0.046). Linear models outperformed more complex algorithms (neural networks, ensemble methods). Conclusion. Linear models, particularly logistic regression, exhibit high accuracy and reliability in predicting AIC using combined biomarkers and echocardiographic diastolic function parameters. These models show potential for clinical implementation in risk stratification and timely initiation of cardioprotective therapy. Further validation across multi-center patient cohorts is warranted.