Development and multicenter validation of a cfDNA methylation-based PCR assay (myESCAP) with machine-learning prediction for non-invasive detection of esophageal cancer

INTRODUCTION: Esophageal cancer (EC) remains a significant public health burden in China. Current diagnostics modalities such as endoscopy and imaging are invasive, poorly accessible, or insensitive for early-stage cases. DNA methylation alterations in circulating free DNA (cfDNA) are promising for non-invasive cancer detection. OBJECTIVES: We aimed to develop and validate a methylation PCR-based assay for EC detection using EC-specific DNA methylation targets. Here, we report myESCAP, a PCR-based assay that amplifies and quantifies levels of EC-specific methylation in cfDNA METHODS: We conducted a multi-center, retrospective case-control study with 489 EC cases and 555 non-malignant controls (benign esophageal/gastric diseases or asymptomatic) to develop myESCAP, a methylation-specific PCR assay containing 11 EC-specific methylation markers from our previous genome-wide methylation screening, selected for strong discriminative performance and PCR compatibility. cfDNA was collected, bisulfite-converted, and tested by myESCAP, and results were analyzed by six modeling strategies (generalized linear model, k-nearest neighbors, C5.0 decision tree, gradient boosting machine, random forest, support vector machine) to train EC-prediction models using 347 EC cases and 349 non-malignant controls, and validated in 142 EC cases and 206 controls. RESULTS: . The six models showed robust performance in the training cohort (average AUC = 0.835, average sensitivity = 76.13%, average specificity = 90.88%). They maintained performance in the validation cohort (average AUC = 0.823, average sensitivity = 75.59%, average specificity = 89%). Specifically, for stage I cases, myESCAP had an average sensitivity of 47.2%, highlighting the challenge of early detection with limited circulating tumor DNA and the need for further optimization. CONCLUSION: myESCAP accurately differentiated EC from non-malignant controls. The six models collectively confirm myESCAP's robustness and clinical potential as a cost-effective, non-invasive diagnostic tool for EC.