Position: AI-Driven Risk Stratification is Essential for Affordable Early Detection of Cancer

Asif Khan, Daniel Ritter, Duncan Terell Forster, Moshir Harsh, Chunlei Zheng, Abdullah Kuziez, Rose Orenbuch, Artem Gazizov, Debora Susan Marks, Nathanael R. Fillmore, Chris Sander

Keywords: Cancer risk, early detection of cancer, AI for medicine, GenAI for healthcare

Abstract: Early detection of cancer offers the best chance for a cure, yet population-wide screening is currently not feasible or not sufficiently implemented for most malignancies due to a number of factors. These include heterogeneity of risk, limited technology for early detection, rarity of cases, lack of funding and lack of societal acceptance. This position paper argues that AI-driven risk stratification using longitudinal, population-scale electronic health records (EHRs) provides a practical first step toward scalable surveillance programs. We outline a three stage program of \emph{prediction--detection--intervention}, in which AI-based risk models nominate small high-risk cohorts, who can be given advanced tests, benefit from emerging early detection technologies and timely therapy or preventive care. We emphasize the importance of model evaluation metrics such as PPV and SIR, which inform real-world feasibility, cost effectiveness, and alignment with healthcare system capacity. As an example, we trained a pancreatic cancer risk model on the US-VA database of 15.9M patients, concluding that focusing on high-risk patients can meaningfully support a realistic surveillance program. We argue that AI-driven risk stratification with proper deployment as decision support, integration into clinical early detection, and intervention workflows, and coordinated efforts across research and healthcare systems have the potential to transform cancer care.

Submission Number: 90