Multimodal Artificial Intelligence System for Risk-Adapted Cancer Survivorship Surveillance: A Multicenter Target Trial Emulation.

Abstract

The growing population of cancer survivors faces immense monitoring burdens due to rigid follow-up guidelines, such as the intensive surveillance schedules recommended by the National Comprehensive Cancer Network (NCCN). To address this issue, we engineered a multimodal artificial intelligence (AI)-based decision support system that integrates biological domain data (magnetic resonance imaging) and physical treatment domain data (radiotherapy dose maps) to guide individualized care. Using stage II nasopharyngeal carcinoma (N=2,148 across five centers) as a model, we first implemented a target trial emulation framework to confirm the safety of treatment de-intensification and establish a baseline for streamlined surveillance. We then trained a Transformer architecture to predict individualized treatment failure timing and translated these predictions into a risk-adapted surveillance strategy. In the target trial emulation, omitting concurrent chemotherapy demonstrated comparable survival outcomes to concurrent chemoradiotherapy across all cohorts, establishing a safely de-intensified clinical baseline. Subsequently, the AI system achieved high-fidelity predictions, with an area under the curve of 0.991 internally and 0.986 in the multi-institutional external validation cohort. This AI-guided strategy substantially reduced the need for follow-up visits for over 90% of failure-free patients, while recommending a maximum of only six visits for high-risk individuals over a five-year period, demonstrating a high sensitivity for detecting true failures. This generalizable AI framework can seamlessly complement the current NCCN guidelines, offering a transformative, data-driven solution that reduces the global monitoring burden of cancer survivorship care.