Development of a Pan-Cancer Screening Model Based on Blood Biomarkers

Neoplasms

pan-cancer, blood biomarkers, machine learning, early detection

This study aims to develop a pan-cancer screening model using routine blood biomarkers (including complete blood count, biochemical tests, coagulation panel, and tumor markers). The study is retrospective, collecting data from approximately 10,000,000 cancer patients diagnosed at multiple centers in China between January 2006 and September 2025. All patients have confirmed pathological diagnosis and complete blood test records. A Mixture of Experts (MoE) machine learning model will be built to predict the presence of various cancers (e.g., gastric, colorectal, liver, lung, ovarian cancer). The goal is to establish a low-cost, non-invasive screening tool suitable for large-scale population screening.

Background: Cancer is a leading cause of death worldwide. Early detection improves survival, but current screening methods (e.g., endoscopy, imaging) are invasive, costly, or not widely accessible. Blood-based biomarkers offer a non-invasive, repeatable, and cost-effective alternative. Objective: Primary: To establish a pan-cancer screening model based on blood biomarkers. Secondary: To combine multiple blood markers for identifying high-risk populations. Exploratory: To develop a cost-effective, scalable screening technology. Study Design: This is a multicenter, retrospective study. Data will be collected from 15 participating hospitals in China, including Zhejiang Cancer Hospital, Tongling People's Hospital, Pingyang People's Hospital, Fenghua People's Hospital, Shaoxing Central Hospital, Bingqi General Hospital, the Second Affiliated Hospital of Jiaxing University, Taizhou Hospital of Zhejiang Province affiliated to Wenzhou Medical University, Affiliated Jinhua Hospital, Zhejiang University School of Medicine, Yunnan Cancer Hospital, Xianju People's Hospital, Longhua Hospital, Shanghai University of Traditional Chinese Medicine, No.9 Hospital Ningbo, Norinco General Hospital, Quzhou Kecheng People's Hospital. Participants: Approximately 10,000,000 patients aged 18-80 years with pathologically confirmed cancer (including respiratory, digestive, urogenital, nervous, endocrine, and soft tissue malignancies). Exclusion criteria: presence of non-cancer organic diseases, hematologic disorders, immunodeficiency (e.g., AIDS), or incomplete data. Data collection: Blood biomarkers including complete blood count, biochemical tests (liver/kidney function, glucose, lipids), coagulation (PT, APTT, TT, fibrinogen), and tumor markers (e.g., CEA, CA19-9, AFP, CA125, etc.) along with clinical data (age, sex, height, weight, diagnosis) will be extracted from medical records. Statistical analysis: A Mixture of Experts (MoE) architecture with deep residual networks, attention-based gating, and feature interaction (FM + deep neural networks) will be used. Multi-task learning, Focal Loss for class imbalance, and adaptive sample weighting will be applied. Model performance will be evaluated for sensitivity, specificity, and AUC. Ethics: Approved by the Ethics Committee of Zhejiang Cancer Hospital (IRB-2025-1319\[IIT\]). Because this is a retrospective study using de-identified data, the committee approved a waiver of informed consent for patients without prior general consent, in accordance with Chinese regulations and the Declaration of Helsinki. Data will be encrypted and stored securely for 15 years after study completion. Dissemination: Results will be published in peer-reviewed journals and presented at conferences.

No related papers found yet.

China