The Clinical Need Early-stage non-small cell lung cancer (NSCLC) is treated with surgical resection, but approximately 30-50% of patients experience recurrence within five years. Current staging systems cannot reliably identify which stage I-IIIA patients need intensified surveillance or adjuvant therapy, leading to both overtreatment and missed recurrences.
RADAR CARE Architecture The RADAR CARE system is a multimodal transformer-based deep learning model that integrates three data streams: 64 clinical, pathological, and molecular baseline factors; longitudinal CT scan data encoded using ClinicalBERT-style embeddings (768-dimensional vectors); and serial laboratory values (52-dimensional). The model produces a RADAR score - a real-time, continuously updated recurrence risk estimate.
Study Scale Published in JCO Precision Oncology in 2025, this study from Samsung Medical Center in South Korea enrolled 14,177 patients with stage I-IIIA NSCLC diagnosed between January 2008 and September 2022 - one of the largest single-center NSCLC cohorts ever assembled for a deep learning study.
Clinical-Molecular Input Module Sixty-four baseline features spanning demographics, tumor pathology, TNM stage, driver mutation status (EGFR, ALK, KRAS), PD-L1 expression, and surgical approach were encoded as a structured input vector. Feature importance analysis later identified stage, surgical margin, and driver mutation status as the highest-weight predictors.
CT Imaging Encoding Rather than processing raw CT pixels, the team encoded radiology reports and structured CT findings using a ClinicalBERT language model that produces 768-dimensional contextual embeddings. This approach captures nuanced language patterns in imaging reports - such as descriptions of lymph node size, ground-glass opacity extent, and vascular invasion - that contain prognostic information beyond simple measurements.
Longitudinal Laboratory Integration Serial laboratory values across 52 parameters - including complete blood count, liver function tests, inflammatory markers, and tumor markers such as CEA - were encoded as time-series inputs. The transformer architecture handles variable-length longitudinal sequences via attention mechanisms, allowing the model to weight more informative time points more heavily.
Overall Model Performance The RADAR CARE model achieved an overall AUC of 0.854 for recurrence prediction across all stages. Stage-stratified performance was strongest in stage I (AUC 0.872) and decreased for stage II (AUC 0.737) and stage III (AUC 0.724), reflecting the greater biological heterogeneity and complex treatment interactions in later stages.
Four RADAR Score Trajectory Patterns Longitudinal analysis of RADAR scores over post-surgical follow-up identified four distinct risk trajectory patterns: persistently low risk, initially low then rising, persistently high risk, and declining risk. Patients with rising RADAR trajectories had significantly higher recurrence rates even when their initial post-surgical scores appeared favorable, enabling earlier identification of late-developing recurrences.
Clinical Utility for Adjuvant Treatment Patients with high RADAR scores who received adjuvant chemotherapy or targeted therapy showed significantly better outcomes than high-risk patients who did not, while low-RADAR patients showed no meaningful benefit from adjuvant treatment. This suggests the RADAR score could be used prospectively to select which patients should receive adjuvant therapy after resection.
Handling Variable-Length Time Series Traditional recurrence prediction models use a single snapshot of patient data at one time point. The RADAR CARE transformer processes months to years of follow-up data as a sequence, allowing it to detect subtle trends in lab values or imaging that precede clinically obvious recurrence by weeks to months.
Attention Mechanism Interpretability The transformer's attention weights provide a degree of interpretability: the model highlights which time points and data modalities were most informative for each patient's risk score. Clinical review of attention maps revealed that rising CEA trajectories, increasing lymph node size on CT reports, and specific inflammatory marker patterns were consistently attended to before recurrence events.
ClinicalBERT vs. Raw CT Processing Using NLP-encoded radiology reports rather than raw CT images reduces computational costs dramatically and allows the model to leverage the clinical expertise embedded in radiologist language. This pragmatic design choice also makes the model compatible with standard clinical imaging workflows without requiring specialized GPU infrastructure for pixel-level CNN processing at scale.
Against TNM Staging TNM staging classifies patients by anatomic disease extent but ignores molecular markers, treatment response signals, and longitudinal physiological changes. The RADAR CARE model substantially outperformed TNM stage alone in identifying patients who developed recurrence within each stage category, demonstrating the additional value of integrating multi-dimensional data.
Against Existing Biomarker Models Previously published single-biomarker or limited-feature models for NSCLC recurrence prediction typically achieve AUCs in the range of 0.70-0.80. RADAR CARE's 0.854 represents a clinically meaningful improvement, particularly at the high-sensitivity operating points relevant for surveillance planning.
Real-Time Updating Advantage Unlike static prognostic models, RADAR CARE updates its risk score at each follow-up visit as new lab and imaging data are added. This real-time updating is a fundamentally different paradigm from standard oncology risk tools and aligns with the goal of adaptive, personalized post-surgical monitoring.
Single-Center Training Limitation Although 14,177 patients provides a large training set, the model was trained and tested at Samsung Medical Center with its specific patient demographics, surgical practices, and imaging protocols. External validation across Western and other Asian populations is essential before international deployment.
ClinicalBERT Report Dependency The model relies on standardized radiology report language. Hospitals with non-standardized, abbreviated, or non-English reporting practices may see degraded performance. Adapting the NLP component to diverse linguistic and reporting styles is a necessary next step.
Prospective Clinical Trial Integration The highest priority for RADAR CARE is a prospective randomized trial in which high-risk RADAR score patients are randomized to RADAR-guided adjuvant therapy versus standard of care. Only such a trial can establish whether using the RADAR score to guide treatment decisions translates to improved survival outcomes, the ultimate standard for clinical adoption.