AI Tools Reshape Post-Surgical Cancer Surveillance

A new class of artificial intelligence models is prompting healthcare leaders to reconsider the long-standing assumptions behind post-surgical cancer surveillance. The recent deployment of the RADAR CARE system by Samsung Medical Center offers a case in point. Trained on more than 14,000 early-stage non-small cell lung cancer (NSCLC) cases, the transformer-based model provides individualized predictions of recurrence risk within the first year after surgery.

The tool, which synthesizes clinical, imaging, pathology, and patient history data, assigns risk scores that inform surveillance intensity. These scores categorize patients into low, intermediate, or high-risk groups. The result is a stratified pathway for ongoing care, departing from the traditional model of uniform follow-up schedules. It is a development with clear operational, clinical, and financial consequences.

Stage-Based Protocols Are Becoming Operational Liabilities

Historical cancer staging systems have shaped surveillance and treatment planning for decades. However, Samsung’s research found that risk scores derived from the RADAR CARE model were often more predictive of recurrence than stage classification alone. For instance, early-stage patients with high AI-assigned scores were more likely to experience recurrence than some later-stage patients with lower scores.

This finding reinforces a growing consensus across clinical research that anatomical staging alone cannot capture the full variability of tumor behavior. A 2023 study published in JAMA Oncology reached similar conclusions, noting that biological and genomic factors often override traditional stage-based assumptions. Uniform follow-up protocols based solely on stage may therefore lead to both under- and over-monitoring, placing unnecessary strain on patients and health systems alike.

For clinical leadership, this presents a challenge. Systems must now determine whether legacy surveillance protocols can be reconciled with AI-driven precision risk scores or whether wholesale redesign is required.

Surveillance Realignment Demands Systemic Readiness

The RADAR CARE model showed clear differentiation in recurrence probabilities: high-risk patients were nearly 10 times more likely to relapse or die within a year compared to low-risk patients, regardless of stage. This granularity could enable more targeted post-operative strategies. High-risk patients may benefit from increased imaging frequency or adjunctive therapies, while low-risk patients may avoid unnecessary interventions.

Yet operationalizing this insight requires system-wide alignment. Institutions must assess whether their care coordination infrastructure can support variable surveillance schedules. As noted in a recent Health Affairs analysis, the gap between predictive capability and actionable integration remains one of the core barriers to clinical AI adoption.

Resource allocation, staffing workflows, and care navigation protocols are not easily adapted. A risk-tiered surveillance model introduces complexity that must be absorbed without introducing delay, confusion, or liability exposure.

Economic and Reimbursement Structures Are Misaligned

Despite its clinical potential, predictive surveillance remains constrained by economic realities. Payer systems are not typically designed to accommodate differentiated follow-up schedules, and reimbursement codes may not reflect risk-based modifications. Without financial incentives or billing structures that support AI-informed care pathways, adoption may stall.

Nevertheless, emerging international models provide useful benchmarks. In Singapore, an AI-driven tool for hepatocellular carcinoma recurrence prediction has demonstrated 82 percent accuracy. Early projections suggest that incorporating such models could reduce unnecessary imaging while improving time-to-treatment metrics for high-risk patients.

Similarly, a 2024 study in the New England Journal of Medicine evaluating an AI model for breast cancer recurrence found that tailored follow-up protocols resulted in fewer emergency readmissions and lower diagnostic delays. These early findings underscore the economic and clinical upside of personalized surveillance, provided the broader system can support it.

Healthcare organizations that proceed without reevaluating their reimbursement models may find themselves absorbing the operational costs of AI deployment without capturing its strategic benefits.

Regulatory Friction and Data Localization Questions Persist

The introduction of any AI-driven clinical tool also reactivates a familiar set of regulatory concerns. The U.S. Food and Drug Administration (FDA) has begun to clarify its approach to software as a medical device, including predictive tools. But ambiguity remains around oversight expectations, especially for tools that inform but do not dictate clinical decisions.

Questions also remain around the generalizability of models trained on regionally specific datasets. RADAR CARE was developed exclusively using South Korean patient data from Samsung Medical Center. Its applicability across broader demographic or geographic populations is unproven. For U.S.-based systems considering adoption, model validation must include localized data inputs and real-world testing protocols to ensure performance equity.

Moreover, liability frameworks have not caught up to this new paradigm. Should a patient relapse under a low-risk designation, providers may face legal scrutiny regarding their use—or non-use—of available AI tools. Until regulatory and legal frameworks mature, many systems may choose caution over innovation.

A Shift Toward Risk-Informed Precision Management

The use of RADAR CARE is part of a larger global pivot toward integrating AI into precision oncology workflows. In partnership with AstraZeneca, Lunit has accelerated genomic testing for NSCLC using AI to streamline diagnostic sequencing. These developments suggest that risk prediction models are not isolated tools but components of a larger ecosystem where early identification, genetic profiling, and individualized planning coalesce.

What remains unresolved is how quickly health systems can adjust their processes to accommodate these advances. Clinical decision support systems, staffing models, and payer contracts must all evolve in parallel. Otherwise, predictive capabilities may remain siloed—technologically advanced but strategically inert.

The transition away from stage-based surveillance toward AI-informed management is no longer speculative. But realizing its full benefit requires readiness across the continuum of care. Risk, once treated as a static attribute, is now a dynamic forecast. The systems that adapt to this reality will lead the next era of oncologic care delivery..