Abstract 596P
Background
Existing surveillance protocols for early-stage non-small cell lung cancer (NSCLC) fail to consider individualized risk factors for recurrence. Treatment strategies are primarily determined based on the TNM staging system, which classifies cancer based on tumor size and metastasis. This study aimed to develop and validate a deep-learning model that leverages comprehensive clinical data to predict recurrence in real-time.
Methods
This study included patients with histologically confirmed NSCLC (stages I-III) who underwent curative surgery between January 2008 and September 2022. We developed a transformer-based multimodal deep-learning model. The model integrated baseline clinical and pathological data (64 factors), longitudinal laboratory data (52 factors), and routine practice imaging data (768 factors) from CT, PET, MRI, and bone scans. The primary objective was to predict the likelihood of recurrence within one year from the monitoring point. Model performance was evaluated using the area under the curve (AUC), sensitivity, and specificity.
Results
A total of 14,177 patients were enrolled, with 10,262 in stage I, 2,380 in stage II, and 1,703 in stage III. During surveillance, 177,246 radiologic interpretation records were utilized. Baseline RADAR scores were 0.324 (SD 0.256) for stage I, 0.660 (SD 0.210) for stage II, and 0.824 (SD 0.140) for stage III. The model achieved an overall AUC of 0.854, with a sensitivity of 86.0% and specificity of 71.3% for predicting one-year recurrence across all stages. Specifically, the AUC was 0.872 for stage I (sensitivity 83.6%, specificity 77.8%), 0.737 for stage II, and 0.724 for stage III. In the total population, the odds ratio of relapse occurring within one year of the relevant follow-up time increases by 1.27 times for each 0.1 increase in the RADAR score (95% CI, 1.24–1.29, P-value < .0001).
Conclusions
This study introduces an advanced deep-learning model that utilizes multimodal data from routine clinical practice to predict recurrence in early-stage NSCLC. The RADAR score provides timely risk assessments for clinicians, potentially guiding risk-adapted surveillance and aggressive adjuvant treatment strategies.
Clinical trial identification
Editorial acknowledgement
Legal entity responsible for the study
The authors.
Funding
Korea Association for ICT Promotion AI Voucher.
Disclosure
All authors have declared no conflicts of interest.