Advanced ensemble staking model employing cfDNA fragmentation for early detection of esophageal and gastric cancer

Abstract

Esophageal and gastric cancers are aggressive malignancies with poor prognoses due to late-stage diagnosis. Our study recruited 275 healthy participants, 201 gastric cancer patients, 74 esophageal patients and 103 patients with precancerous conditions. The participants were assigned into training and validation cohorts. After processing a low-depth whole genome sequencing for all plasma samples, a stacked ensembled model was constructed, integrating three cfDNA fragmentomic features: Copy Number Variation, Fragment Size Profile, and Fragment Based Methylation. The multi-dimensional model was trained with 5-fold cross-validation, and its performance was evaluated through validation. The detection sensitivity and specificity were validated at 95 % specificity of training set. The stacked ensemble model achieved an AUC of 0.967 in the validation dataset. At a 95 % specificity threshold, the model attained a high sensitivity of 79.2 %, underscoring its clinical utility in distinguishing cancer from healthy individuals. Notably, it achieved sensitivity of 77.4 % and 68.3 % for stage I cases in training and validation cohorts, respectively. The model also identified precancerous conditions effectively, with an AUC of 0.828 and sensitivity of 53.8 % and 71.4 % for gastric and esophageal precancer lesions, while maintaining clear score distinctions in specifying benign diseases. Overall, our stacked model achieved high sensitivity in identifying esophageal and gastric cancer, offering a strong, non-invasive alternative to endoscopy. This approach supports timely intervention and improved patient outcomes by enabling earlier and more targeted treatment.