AI-powered risk score predicts complications in cancer therapy

A group of European researchers has introduced a machine learning–based clinical decision support tool designed to predict complications in cancer patients receiving bevacizumab and its biosimilars. The team developed and validated a model that could transform how oncologists approach personalized treatment planning.

Published in Diagnostics, the study “Using Artificial Intelligence to Develop Clinical Decision Support Systems: The Evolving Road of Personalized Oncologic Therapy” marks a step forward in integrating artificial intelligence (AI) into real-world oncology practice. By combining routine clinical data with advanced algorithms, the researchers sought to provide a practical tool that clinicians can use before treatment decisions are finalized.

How was the AI model built and tested?

The research team developed the model using data from 395 treatment episodes of patients receiving bevacizumab, a monoclonal antibody therapy widely applied in solid tumors. The study design was prospective and real-world, focusing exclusively on pre-treatment variables such as demographics, medical history, tumor characteristics, and laboratory results. This ensured that the model avoided bias from post-treatment data leakage, which can distort predictive performance.

Several machine learning approaches were compared, including logistic regression, Random Forest, and XGBoost. After optimization and evaluation, the Random Forest model performed best on the held-out test dataset, delivering an accuracy of 70.63 percent, sensitivity of 66.67 percent, specificity of 73.85 percent, and an AUC-ROC of 0.75. These results demonstrated that machine learning could provide robust, clinically relevant predictions in oncology risk stratification.

To translate the model into something immediately usable in clinics, the researchers developed a logistic regression–based risk score derived from the most predictive features identified during analysis. The risk score achieved AUC-ROC 0.720, with balanced performance across accuracy, sensitivity, specificity, and F1 score.

What risks can the AI tool predict for cancer patients?

The key outcome predicted was the likelihood of experiencing bevacizumab-related complications, which can include hypertension, bleeding, or thromboembolic events. By identifying patients at higher risk before treatment begins, clinicians can adjust monitoring, modify therapeutic approaches, or counsel patients more effectively.

Key predictors included age above 65, anemia, elevated urea levels, leukocytosis, tumor differentiation status, and cancer stage. Each of these factors contributed to risk estimation in a multi-factorial way, underlining how complications arise not from a single driver but from a combination of interacting patient characteristics.

The model’s interpretability was enhanced using SHAP analysis, which confirmed that age and diagnosis category were among the most influential features. Rather than providing opaque predictions, the tool highlights why certain patients are categorized as higher risk, making it more suitable for clinician trust and adoption.

Importantly, the model was not intended to replace medical judgment but to support oncologists in making evidence-informed decisions. In practice, a patient flagged as high-risk could receive more intensive follow-up and tailored interventions, while those at lower risk might avoid unnecessary procedures or over-monitoring.

How can this tool shape the future of personalized oncology?

The researchers noted that the final product is not just an abstract model but a ready-to-use offline HTML calculator. This form-based tool categorizes patients into low, intermediate, and high risk, allowing point-of-care use without the need for constant internet connectivity or specialized infrastructure. Such accessibility is particularly important in resource-constrained settings, where advanced predictive analytics are rarely available.

From a clinical perspective, the tool is best suited for rule-in strategies. Internal validation showed that while sensitivity was moderate, specificity was higher, making the system more reliable for confirming which patients require extra vigilance. This reflects a cautious but meaningful role for AI in oncology: enhancing safety by flagging those most likely to face treatment-related complications.

However, the authors acknowledge limitations. The dataset was drawn from a single center, raising questions about generalizability. Class imbalance also reduced sensitivity, and the absence of advanced biomarkers limited the richness of the model. As such, the team calls for multi-center validation, larger datasets, and adaptive learning frameworks that can update the model as new patient data become available.

For future, this research illustrates the broader evolution of personalized oncologic therapy. As treatment regimens become increasingly complex, AI tools like this can provide a critical layer of support. They promise not only more efficient decision-making but also improved outcomes for patients who might otherwise be exposed to preventable risks.