2606 - Machine Learning-Based Prognostic Clinical-Genomic Model of Disease Recurrence after Chemoradiation and Durvalumab in Unresectable Stage II-III NSCLC
Presenter(s)
K. Shin1, L. Martinka2,3, K. Olabode2,3, Simran2,4, J. Gray5, S. Puri5, A. N. Saltos5, T. Tanvetyanon5, B. Creelan5, A. Chiappori5, C. Lu5, M. Shafique5, K. A. Ahmed6, K. Yamoah2, P. C. Rodriguez7, S. A. Rosenberg2, S. K. Jabbour8, T. J. Dilling2, and J. Kim7,9; 1Division of Radiation Oncology, The University of Texas MD Anderson Cancer Center, Houston, TX, 2Department of Radiation Oncology, H. Lee Moffitt Cancer Center and Research Institute, Tampa, FL, 3Morsani College of Medicine, University of South Florida, Tampa, FL, 4Department of Radiation Oncology, All India Institute of Medical Sciences (AIIMS), New Delhi, India, 5Department of Thoracic Oncology, H. Lee Moffitt Cancer Center and Research Institute, Tampa, FL, 6H. Lee Moffitt Cancer Center and Research Institute, Department of Radiation Oncology, Tampa, FL, 7Department of Immunology, H. Lee Moffitt Cancer Center and Research Institute, Tampa, FL, 8Department of Radiation Oncology, Rutgers Cancer Institute, New Brunswick, NJ, 9Department of Radiation Oncology, Moffitt Cancer Center, Tampa, FL
Purpose/Objective(s): Definitive concurrent chemoradiotherapy (CRT) followed by durvalumab is the standard of care for unresectable locally advanced non-small cell lung cancer (LA-NSCLC), yet many patients still experience relapse. Prognosis traditionally relies on the AJCC clinical stage, which may not capture individual tumor biology. We aim to develop a combined clinical–genomic risk score to identify patients at higher risk of recurrence after CRT and durvalumab.
Materials/Methods: We identified patients with LA-NSCLC treated with definitive CRT followed by durvalumab between 2015 and 2024. Clinical and genomic data were extracted from the institutional database. Patients lacking baseline next-generation sequencing (NGS) were excluded. Gene mutations identified among the 252-gene panel were analyzed as binary variables (mutant: 1 vs. wild-type: 0). The primary objective was variable selection to predict recurrence-free survival (RFS) and overall survival (OS) using LASSO regression with stability selection (cutoff = 0.6, q=25). Recurrence was defined as locoregional and/or distant recurrence. Secondary objectives included estimating hazard ratios (HRs) and constructing a risk score via a Cox proportional hazards model. Finally, the predictive performance of the risk score was compared against the AJCC clinical stage by an ANOVA test.
Results: A total of 114 patients (stage II: 12; stage III: 102) were included, with a median follow-up of 43.0 months (range, 4.5 –85.4). Median age was 68 years (39 –87), and 67 (58.8%) patients progressed. Four variables were selected for RFS: Stage II (vs III) (HR 0.25, 95% CI 0.06–1.01), KMT2A mutation (HR 3.03, 95% CI 1.21–7.58), STK11 mutation (HR 2.55, 95% CI 1.36–4.78), and FLT1 mutation (HR 0.27, 95% CI 0.08–0.88). No variables could predict OS with the determined cutoff. The risk score for RFS was defined as (-1.40 * Stage II) + (1.11 * KMT2A) + (0.94 * STK11) + (-1.31 * FLT1). Using the median risk score as a cutoff, patients were divided into a low-risk group (n=20) and a high-risk group (n=94). Median RFS was 10.8 months in the high-risk group versus 27.0 months in the low-risk group (p<0.01). When stratified by the AJCC stage alone, Stage II patients did not reach median RFS, whereas Stage III patients had a median RFS of 21.6 months. The risk score outperformed the AJCC clinical stage in predicting RFS (C-index 0.63 vs. 0.54; p < 0.01).
Conclusion: We developed a novel four-factor risk score combining the AJCC stage and three gene mutations (KMT2A, STK11, FLT1) that significantly stratified RFS in stage II–III NSCLC after CRT and durvalumab. The model outperformed the AJCC stage alone and identified patients at higher risk of progression. This suggests that a combined clinical–genomic signature has a role in risk stratification and guiding personalized treatment or follow-up. Prospective external validation is warranted.