Background/Objectives: The use of artificial intelligence (AI) in oncology has the potential to improve decision making, particularly in managing the risk associated with targeted therapies. This study aimed to develop and validate a machine learning-based clinical decision support system (CDSS) capable of predicting complications associated with Bevacizumab or its biosimilars and to translate the resulting predictive model into a clinically applicable tool. Methods: A prospective observational study was conducted on 395 records from patients treated with Bevacizumab or biosimilars for solid tumors. Pretherapeutic variables, such as demographic data, medical history, tumor characteristics and laboratory findings, were retrieved from medical records. Several machine learning models (logistic regression, Random Forest, XGBoost) were trained using 70/30 and 80/20 data splits. Their predictive performances were compared using accuracy, AUC-ROC, sensitivity, specificity, F1-scores and error rate. The best-performing model was used to derive a logistic-based risk score, which was further implemented as an interactive HTML form. Results: The optimized Random Forest model trained on the 80/20 split demonstrated the best balance between accuracy (70.63%), sensitivity (66.67%), specificity (73.85%), and AUC-ROC (0.75). The derived logistic risk score showed good performance (AUC-ROC = 0.720) and calibration. It identified variables, such as age ≥ 65, anemia, elevated urea, leukocytosis, tumor differentiation, and stage, as significant predictors of complications. The final tool provides clinicians with an easy-to-use, offline form that estimates individual risk levels and stratifies patients into low-, intermediate-, or high-risk categories. Conclusions: This study offers a proof of concept for developing AI-supported predictive tools in oncology using real-world data. The resulting logistic risk score and interactive form can assist clinicians in tailoring therapeutic decisions for patients receiving targeted therapies, enhancing the personalization of care without replacing clinical judgment.
Chitoran et al. (2025) studied this question.