Deep-learning model for real-time prediction of recurrence in early-stage non-small cell lung cancer: A multi-modal approach.

8029 Background: The surveillance protocol for early-stage non-small cell lung cancer (NSCLC) is not contingent upon individualized risk factors for recurrence. Longitudinal risk calculation for the prediction of recurrence was based on circulating tumor DNA. This study aimed to develop a deep-learning model using comprehensive data from clinical practice for practical longitudinal monitoring. Methods: A multi-modal deep-learning model employing transformers was developed for real-time recurrence prediction using clinical, pathological, and molecular data at baseline, alongside longitudinal laboratory, and radiologic data during surveillance. Patients with histologically confirmed NSCLC (stage I-III) undergoing curative intent surgery between January 2020 and September 2022 were included. The collected data was divided into training, validation, and test sets in a 6:2:2 ratio. The primary outcome focused on predicting likelihood of recurrence within one year from the monitoring point. This study demonstrated the timely provision of risk scores (RADAR score), determined thresholds, and the corresponding Area Under the Curve (AUC). Results: A total of 14,345 patients were enrolled (10,262 with stage I, 2,380 with stage II, and 1,703 with stage III). The training, validation, and test sets comprised 8,578, 2,866, and 2,901 patients, respectively. The model incorporated 64 clinical-pathological-molecular factors at baseline, alongside longitudinal laboratory, and radiologic text data. Radiologic data of 177,246 was used (mean 12.4 chest CT scan per patient) during surveillance. Baseline RADAR score was 0.20 (standard deviation Std 0.17) in stage I, 0.51 (Std 0.20) in stage II, and 0.64 (Std 0.19) in stage III. The AUC for predicting relapse within one year from that monitoring point was 0.847 across all stages with the sensitivity of 80.6% and specificity of 74.4% (Table) (AUC=0.851 in stage I, AUC=0.763 in stage II, and AUC=0.721 in stage III). Conclusions: This pilot study introduces a deep-learning model utilizing multi-modal data from routine clinical practice for predicting relapse in early-stage NSCLC. It demonstrates the timely provision of risk score, RADAR score, to clinicians for recurrence prediction, potentially guiding risk- adapted surveillance strategies and aggressive adjuvant systemic treatment. Notably, the RADAR score exhibited a high prediction ability for relapse in stage I disease. Table: see text