基于常规实验室检测的AI赋能癌症风险评估模型用于富集高癌症风险个体

Background : Early cancer detection improves survival, yet population-wide screening remains costly and logistically challenging. Traditional cancer risk models, such as lifestyle-based assessments and polygenic risk scores, offer limited sensitivity. In contrast, routine laboratory tests (complete blood count, urinalysis, and biochemical panels) are already widely performed in primary care and may harbor latent cancer signals. We developed an AI-based cancer risk assessment (CRA) model using these existing data to enrich cancer cases and improve the cost-effectiveness of downstream multi-cancer early detection (MCED) tests. Methods : Routine laboratory data from 5,376 individuals (1,399 cancer; 3,977 non-cancer) across two hospitals were retrospectively collected and split into training and validation cohorts. A total of 43 features were selected to calculate CRA scores using a gradient boosting framework. All participants also underwent OncoSeek testing, a validated MCED assay combining multiple protein tumor markers with clinical data, applied to CRA-positive individuals. Results: The CRA model achieved AUCs of 0.738 and 0.802, with sensitivities of 87.5% and 90.0% at specificities of 50.0% and 47.2% in the training and validation cohorts. Among individuals predicted positive (CRA > 0.4, n = 3241), cancer prevalence increased from 26.0% to 38.2% (1.5-fold enrichment). In this enriched group, OncoSeek achieved 37.0% sensitivity at 97.3% specificity, and 56.6% sensitivity when restricted to its covered 14 cancer types. Furthermore, within the high-risk subgroup (CRA > 0.88, n = 231), 65.4% were cancer patients, with a higher proportion of advanced disease (Stage IV: 42.2%) than the overall cohort (25.0%), indicating potential to flag patients needing prompt diagnostic evaluation.A population-level simulation in 1 million adults ≥50 years (incidence = 0.9%) showed that adding CRA before OncoSeek increased cancer incidence to 1.5% (1.7-fold), reduced false positives by 33.3%, and halved total screening cost ($80.0 M → $40.5 M), lowering cost per detected case by 44.1%, demonstrating improved efficiency and cost-effectiveness. Conclusion : The AI-based CRA model, leveraging routine primary care data, provides substantially higher sensitivity and broader applicability than polygenic risk scores. Beyond enriching cancer incidence for screening, CRA can also flag individuals with strong cancer signals who may require prompt diagnostic evaluation in EMR system. When integrated as a front-end enrichment step before MCED testing, this approach effectively filters out low-risk individuals, reduces unnecessary downstream testing, and halves screening costs with only a modest (~9%) sensitivity reduction. The combined CRA + MCED framework offers a scalable and economically sustainable pathway for precision population-level cancer early detection.