Accurate differentiation of malignant and benign gastric lesions using cell-free DNA biomarkers.

3047 Background: Early detection of gastric cancer is challenging due to the invasive nature of current diagnostic methods and the difficulty in distinguishing cancer from benign gastric conditions. Cell-free DNA (cfDNA) features have emerged as promising biomarkers for non-invasive detection. This study aims to develop and evaluate a machine learning model utilizing cfDNA features for early gastric cancer detection. Methods: We developed an ensemble machine learning model incorporating four cfDNA features: repeat elements, fragment-based methylation, focal copy number variation, and fragment size pattern. The model was trained using cfDNA data from 150 gastric cancer patients and 153 individuals with stomach-related conditions. The ensemble model was validated using a cohort of 149 cancer patients, 149 individuals with high-risk benign lesions, and 50 low-risk benign lesions. Risk is stratified according to the Correa’s Cascade. Results: The ensemble machine learning model developed using four cfDNA features achieved an AUROC of 0.913 in the training cohort and 0.912 in the testing cohort for distinguishing gastric cancer patients from individuals with stomach-related complications, which outperformed individual cfDNA features. A decision threshold of 0.418, established via cross-validation, was set to ensure at least 95% sensitivity in the training cohort. This threshold enabled accurate binary classification in the validation cohort, with model scores correlating with cancer stage and tumor differentiation, supporting its potential for clinical risk stratification. Model scores effectively differentiated cancer from high-risk individuals, with significantly lower scores in non-cancer groups compared to precancerous or Stage I–III cancer cases (Non-cancer group median score = 0.35, precancer group median score = 0.48, cancer group median score = 0.61). Additionally, the model assigned significantly higher cancer prediction scores to gastric cancer cases compared to high-grade intraepithelial neoplasia (p = 0.003). In the validation dataset, sensitivities were 92.9% (95% CI: 85.3%–96.7%) for Stage I, 96.3% (95% CI: 81.7%–99.3%) for Stage II, and 100% (95% CI: 83.2%–100%) for Stage III. Sensitivities for well-, moderate, and poorly differentiated tumors were 91.7%, 92.2%, and 100%, respectively. Of note, specificity for the detection of cancer was 66% in the training cohort and 71% in the validation cohort. Conclusions: Our findings demonstrate the potential of cfDNA-based machine learning models as a non-invasive and accurate diagnostic tool for early gastric cancer detection. By reducing reliance on invasive procedures, this approach could enhance clinical workflow efficiency and improve patient outcomes. Further validation in larger, independent cohorts is needed to support clinical implementation.