Advanced Anomaly Detection Framework for Enhancing Prediction of Severe Health Outcomes in Cancer Patients Undergoing Radiotherapy

The health trajectories and radiation-induced toxicities of cancer patients receiving radiotherapy are shaped by complex interactions among disease progression, treatment protocols, and symptom burden arising from both illness and therapy. Accurate prediction of severe clinical outcomes, such as debilitating pain, depression, and sleep disturbances, is critical to advance precision oncology and personalized care. However, extreme class imbalance in patient-reported outcomes (PROs) datasets, where severe cases are underrepresented, poses a significant challenge for conventional predictive modeling. Using PROs data, this study proposes a novel anomaly detection framework to address this imbalance by systematically evaluating supervised, semi-supervised, and unsupervised machine learning approaches. Our findings reveal that semi-supervised anomaly detection methods, including techniques such as DeepSAD, PReNet, and XGBOD, consistently outperform traditional models in identifying minority-class patients with severe health deterioration. Notably, unsupervised methods such as Autoencoder-based anomaly detection also demonstrate promising versatility by balancing sensitivity to both majority (mild/intermediate) and minority (severe) cohorts. These results underscore the potential of anomaly-driven frameworks to enhance clinical decision making, allowing early intervention for high-risk patients while optimizing resource allocation in oncology practice.