Machine learning and plasma growth differentiation factor-15 (GDF15) for early risk stratification of heart failure progression in a long-latency doxorubicin cardiomyopathy model

BACKGROUND: Anthracycline-induced cardiotoxicity is a major cause of late heart failure (HF) in cancer survivors. Yet early identification of individuals at risk for accelerated cardiotoxic progression remains a clinical challenge. This study leveraged a rat model of long-latency doxorubicin cardiotoxicity to identify early features associated with progression to a prespecified 7-month HF endpoint and to develop a machine learning risk-stratification framework. METHODS: Doxorubicin-treated rats (both sexes, n = 32 doxorubicin treated and n = 22 saline) were monitored with monthly echocardiography and monthly plasma concentrations of growth differentiation factor-15 (GDF15) for up to 12 months. We defined the primary HF endpoint as the occurrence of severe systolic dysfunction (fractional shortening (FS) < 30%) by 7 months post-chemotherapy. A Logistic Regression (LR) model was compared to Random Forest (RF) and XGBoost (XGB) classifiers in distinguishing animals that reached this endpoint by 7 months from those with delayed or absent progression within that time frame. Model performance was evaluated by area under the ROC curve (AUC), Brier score, calibration plots, decision curve analysis (DCA), and DeLong tests for AUC differences. Feature importance was examined via permutation testing, SHapley Additive exPlanations (SHAP) values, and accumulated local effects (ALE) plots. RESULTS: Using 1,000 repeated stratified 60/40 train-validation splits, models were retrained and evaluated on the corresponding held-out validation subsets. LR achieved a median AUC of 0.81, RF 0.97, and XGB 0.93. Pairwise DeLong tests were not statistically significant, so the numerically higher AUCs for RF and XGB should be interpreted as suggestive rather than definitive evidence of superior discrimination. The machine learning models showed lower Brier scores (approximately 0.06 vs. 0.12 for logistic regression) and higher estimated net benefit in decision-curve analysis within this dataset. A label-permutation test indicated that RF performance exceeded chance (p < 0.05). Risk groups defined by machine learning-predicted probability showed strong time-to-HF separation on Kaplan-Meier analysis (log-rank p < 0.0001). DISCUSSION: Key predictors included early elevation in plasma GDF15 levels, early decline in FS (ΔFS), body weight, and sex. In this doxorubicin cardiotoxicity animal model, RF and XGB showed numerically higher discrimination than logistic regression, while sensitivity analyses indicated that GDF15 and ΔFS retained predictive signal beyond sex alone. CONCLUSION: These findings support further study of multi-parameter early monitoring and predictive modeling for risk stratification of accelerated cardiotoxic progression in animal models and, ultimately, in appropriately validated clinical cohorts.