Accounting for Cancer Patients with Severe Outcomes: An Anomaly Detection Perspective

Health outcomes and radiation-induced toxicities for cancer patients undergoing radiation therapy are influenced by several factors, including the disease process, treatment plan, and various symptoms produced by both the disease and treatment. Accurate prediction and assessment of a patient’s health status are pivotal for precision and personalized healthcare, especially for patients suffering severe outcomes such as pain, depression, and sleep disorders. Motivated by the issue of extreme class imbalance, this study investigates a set of unsupervised anomaly detection approaches to classify patients with mild/intermediate and severe outcomes using patient-reported outcomes (PROs) datasets. We found that the HBOS method demonstrated superior performance for the minority class, representing cancer patients with severe health statuses. Moreover, IForest and KNN showcased good potential by effectively considering both majority and minority patients. This study may provide insightful guidelines for clinical practice.