Development and prospective validation of a novel cfDNA-based diagnostic model for the early detection of pancreatic cancer.

Abstract

4191 Background: Pancreatic cancer (PC) is one of the most lethal malignancies, with a 5-year survival rate below 10%, primarily due to late-stage diagnosis. Existing diagnostic markers, like CA19-9, show inadequate sensitivity and specificity for early detection. To address this critical gap in early PC detection, this study developed and prospectively validated a cfDNA-based diagnostic model integrating fragmentomics features, including copy number variations (CNVs), fragment size ratios (FSR), and orientation-aware cfDNA fragmentation (OCF). Methods: A multicenter study was carried out with a case-control cohort (n = 467) for model development and a prospective cohort (n = 1,926) for clinical validation. Plasma cfDNA underwent low-pass whole-genome sequencing to extract fragmentomics features like CNVs, FSR, and OCF. A stacked ensemble machine-learning model was built based on case-control data and validated in the prospective cohort of PC elevated-risk individuals with diabetes or obesity. Sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were evaluated and compared with those of CA19-9. The follow-up was intended to last for 3 to 5 years, with the current follow-up period ranging from 12 to 24 months. Results: In the case-control cohort, the cfDNA-based model achieved AUCs of 0.9799 and 0.9622 in the training and validation sets, respectively, with both sensitivity and specificity exceeding 90%. In the prospective cohort (n = 1,926), for the 8 PC cases identified, the cfDNA model demonstrated a sensitivity of 75%, specificity of 98.1%, and PPV of 15.2% for detecting PC, significantly outperforming CA19-9 (sensitivity: 12.5%, specificity: 94.3%, PPV: 0.9%). Notably, the cfDNA model detected all 3 Stage 0 cases, 1 of 3 Stage I cases, and both Stage II cases, providing a median lead time of 227.5 days (range: 45–298 days) compared to imaging modalities. In contrast, CA19-9 detected only one Stage II case out of eight confirmed PC cases (12.5%). The model demonstrated significant potential in stratifying pancreatic cysts into high-risk and low-risk categories. While CA19-9 is ineffective in detecting either high-risk or benign cysts within the prospective cohort, the cfDNA model successfully differentiates between high-risk and low-risk pancreatic cysts (e.g., high-risk IPMN, 1/1 = 100% sensitivity; low-risk SCN, 0/1 = 0% false positive), which further underscores its clinical utility. Conclusions: This study is the first to validate a cfDNA-based diagnostic model for PC in a large elevated-risk population, showing superior performance and significant lead time benefits. The model detects PC earlier with much higher sensitivity and specificity than CA19-9, promising better outcomes with earlier treatment. The findings highlight cfDNA's potential for non-invasive PC screening in clinical settings.