RADAR CARE Study: A Multicenter Deep Learning Model for Real-Time Prediction of Recurrence in Early-Stage Non–Small Cell Lung Cancer Using a Multimodal Approach

Why It Matters

Accurate, real‑time recurrence prediction enables clinicians to tailor adjuvant therapies, potentially improving survival and reducing overtreatment in early‑stage NSCLC. The model’s robust, cross‑institution performance positions it for broader clinical adoption and informs future AI‑driven oncology tools.

Key Takeaways

• Model predicts 1‑year NSCLC recurrence with 0.874 AUC.
• Trained on 15,439 patients across three Korean hospitals.
• Performance stable across EGFR and ALK mutation subgroups.
• Stage‑I predictions achieve highest AUC of 0.892.

Pulse Analysis

Predicting recurrence in early‑stage non‑small cell lung cancer has long been a clinical challenge, hampered by heterogeneous patient data and limited external validation of AI tools. Recent advances in transformer architectures allow models to ingest diverse data types—clinical notes, pathology reports, lab values, and imaging—simultaneously, creating a richer patient representation than single‑modality approaches. This multimodal strategy aligns with a broader shift in oncology toward precision medicine, where risk stratification informs decisions about adjuvant chemotherapy, radiotherapy, or enrollment in clinical trials.

The RADAR CARE study leveraged a dataset of over 15,000 patients collected from Samsung Medical Center, Kyung Hee University Hospital, and St. Vincent’s Hospital, spanning 2008‑2024. By splitting the cohort into training, validation, and independent test sets, the researchers avoided overfitting and demonstrated genuine external validity. The model’s AUC of 0.874 surpasses many earlier NSCLC recurrence predictors, while stage‑specific performance—0.892 for stage I—highlights its sensitivity in the lowest‑risk group where overtreatment is a concern. Importantly, the algorithm maintained accuracy across EGFR and ALK mutation subgroups, suggesting that genomic variability does not erode its predictive power.

Clinically, a reliable RADAR score could be delivered at the point of care, flagging high‑risk patients for intensified surveillance or early systemic therapy. Health systems stand to benefit from reduced recurrence‑related costs and improved patient outcomes. Future work should explore integration with electronic health records, prospective validation in Western cohorts, and regulatory pathways for AI‑based decision support. As the oncology community embraces data‑driven tools, models that prove both accurate and generalizable—like this one—will shape the next generation of personalized cancer care.