Author: Meixu Chen, Jing Wang 👨🔬
Affiliation: Medical Artificial Intelligence and Automation (MAIA) Lab, Department of Radiation Oncology, UT Southwestern Medical Center, UT Southwestern Medical Center 🌍
Purpose: Cancer treatment outcome prediction plays a pivotal role in guiding therapeutic decisions and optimizing patient care. Traditionally, binary prediction models have been widely used for risk stratification. However, for outcomes of interest that manifest over extended periods, survival prediction (time-to-event) models provide a more mathematically sophisticated, and clinically relevant alternative. This study aims to demonstrate the advantages of survival prediction models over binary models for risk stratification, focusing on robustness and clinical relevance.
Methods: Using the largest publicly available head and neck cancer radiotherapy dataset RadCure (1,800 training, 750 testing samples), we developed both binary and survival prediction models employing CNN, Vision Transformers (ViT), and fused CNN-ViT (CNViT) architectures. The input to the prediction models comprising planning CT, contours of gross tumor volume and involved nodal volume, and patient characteristics, including demographic data, tumor stage data, and treatment related information. Outcomes of interest included patient mortality, cancer local recurrence, regional recurrence, and distant recurrence. Binary models targeted events occurring within 2 years post-treatment, while survival models used risk scores to stratify patients based on validation-derived thresholds. Models were trained on data subsets ranging from 10% to 100% of the training set to evaluate performance under varying data availability.
Results: Survival prediction models consistently outperformed binary models when using more than 40% samples from RadCure training dataset for training. For instance, overall survival models trained on 20% of the data achieved comparable performance to binary models trained on 100%, highlighting their data efficiency and resilience.
Conclusion: Survival prediction models offer significant advantages for stratifying risk in cancer treatment outcomes, especially for delayed events. Their ability to utilize censored data makes them a superior alternative to binary models. We advocate for their broader adoption in research and clinical practice to enhance personalized treatment strategies and improve patient outcomes.