مقالات پذیرفته شده در نهمین کنگره بین المللی زیست پزشکی
AI-Assisted cfDNA Fragmentomics for Ultra-Early Multi-Cancer Detection: A Systematic Review
AI-Assisted cfDNA Fragmentomics for Ultra-Early Multi-Cancer Detection: A Systematic Review
Sana Mahdian Rizi,1,*
1. Affiliation: Students Research Committee, Neyshabur University of Medical Sciences, Neyshabur, Iran
Introduction: Early detection of cancer remains the most effective strategy to improve survival, yet conventional imaging and screening often fail to detect tumors at curable stages. Recent advances in cell-free DNA (cfDNA) fragmentomics—including methylation signatures, nucleosome positioning, and promoter fragmentation entropy (PFE)—combined with artificial intelligence (AI) methods, offer a non-invasive path toward multi-cancer early detection (MCED). Despite rapid progress, evidence on diagnostic performance, tissue-of-origin classification, and clinical readiness remains fragmented and requires systematic synthesis.
Objective:To systematically review and meta-analyse the performance of AI-enabled cfDNA fragmentomic assays for ultra-early detection and tissue-of-origin prediction in multiple cancers.
Methods: We searched PubMed, EMBASE, Web of Science, and Cochrane (Jan 2021–Jul 2025) using terms “fragmentomics,” “cfDNA methylation,” “promoter fragmentation entropy,” “AI/ML,” and “multi-cancer early detection.” Human studies evaluating fragmentomic cfDNA with AI/ML algorithms were included if diagnostic metrics (sensitivity, specificity, AUC) were reported for early-stage cancers. Excluded were animal/in vitro studies, case reports, or non-AI designs.
Two reviewers extracted data on cohort size, cancer type, stage distribution, fragmentomic features, algorithms, validation, and metrics. Risk of bias was assessed with QUADAS-2. Meta-analysis used a random-effects model (DerSimonian-Laird) with heterogeneity (I²) and subgroup analyses by cancer type, feature, and validation design. Software: R (meta, mada), Python (scikit-learn, XGBoost), RevMan for risk of bias.
Results: From 2,173 records, 27 studies (n≈45,300; cancers ≈6,700; controls ≈38,600) met criteria. Ten cancer types were covered, notably breast, lung, colorectal, ovarian, and pancreatic. About half of cases were stage I–II; two cohorts involved asymptomatic screening populations.
Methylation assays (12 studies): Sensitivity 62–88%, specificity 90–97% for early detection.
Fragmentation entropy/nucleosome features (5 studies, e.g. EPIC-Seq, DELFI): AUC 0.93–0.98; tissue-of-origin accuracy 80–90%.
Pooled analysis: Sensitivity 75% (95% CI 70–80%), specificity 94% (92–96%), pooled AUC 0.95 (0.93–0.97) for stage I–II cancers.
Tissue of origin (9 studies): Weighted accuracy 85% (81–89%), decreasing ~5–10% in external cohorts.
Bias/heterogeneity: I² >75% for sensitivity; moderate bias due to limited sample sizes and overfitting risks.
Conclusion: AI-enabled fragmentomic assays demonstrate high accuracy for detecting multiple early-stage cancers and predicting tissue of origin. Technologies such as EPIC-Seq and DELFI are especially promising. However, translation to clinical practice requires standardization of sample processing, external validation in large screening cohorts, inclusion of diverse populations, and cost-effectiveness assessments. With these steps, fragmentomics could redefine cancer prevention and early detection by enabling non-invasive, ultra-early, population-level screening.