BACKGROUND
Cardiovascular toxicities are a recognized complication of immune-checkpoint inhibitors (ICIs), yet systematic cardio-oncology surveillance strategies and integrative data analyses are not well established. The potential of large clinical datasets to inform risk stratification is often underexploited. This study aimed to develop and validate a machine learning model to predict 5-year survival in ICI-treated patients and to demonstrate the utility of a clinician-centric, AI-assisted analytical workflow for leveraging complex registry data.
METHODS
This retrospective, single-center study analyzed data from 515 patients receiving ICI therapy, enrolled in the Essen Cardio-Oncology Registry (EcoR) between February 2018 and March 2023 (mean age 63 ± 14 years; 40% female). The dataset comprised baseline demographic, oncologic, and cardiovascular data, including clinical history, vital signs, electrocardiographic and echocardiographic parameters, laboratory biomarkers, and medication regimens. Entries with excessive missingness were excluded, remaining missing values were interpolated. The cohort was partitioned into training (80%) and testing (20%) sets. A Gradient Boosting Classifier (XGBoost) model was developed in Python to predict survival. Model performance was evaluated using accuracy, precision, recall, F1-score, and the area under the receiver operating characteristic curve (AUROC). Model interpretability was assessed using permutation feature importance and SHAP values.
RESULTS
On the hold-out test set (n = 103), the model demonstrated moderate discriminative ability, yielding an accuracy of 78% and an AUROC of 0.76. The model showed robust performance in identifying 5-year survivors (precision: 0.83, recall: 0.90, F1-score: 0.87), though performance was limited for non-survivors (precision: 0.38, recall: 0.25, F1-score: 0.30), as reflected by the confusion matrix (TN: 75, FP: 8, FN: 15, TP: 5). Feature importance analysis identified non-severe myocardial injury at 6 months, baseline left ventricular ejection fraction (LVEF), immature granulocyte count, baseline hemoglobin, height, and a history of heart failure as the most influential predictors.
CONCLUSION
This study demonstrates the feasibility of implementing a model to predict 5-year survival in patients undergoing ICI therapy with moderate accuracy. Importantly, such explainable models can serve as powerful hypothesis-generating tools by identifying potentially overlooked prognostic markers within complex datasets. This approach exemplifies a practical pathway for integrating machine learning analytics into clinical research. The embedding of such AI-supported tools in cardio-oncology practice holds promise for enhancing data-driven risk stratification and promoting the broader utilization of institutional registry data.
