Multi-tumor biomarker velocity analysis and machine learning integration for early detection of malignant transformation in pulmonary nodules: a Chinese large population-based cohort study

Current clinical strategies for the assessment of pulmonary nodules predominantly rely on static tumor biomarker detection at single time points. This “snapshot” methodology is inherently limited as it frequently overlooks the critical temporal information associated with malignant transformation, often leading to delayed diagnosis for early-stage lung cancer or unnecessary invasive procedures for indeterminate benign nodules. This study aimed to bridge this significant diagnostic gap by developing and validating the biomarker velocity index (BVI), a novel dynamic metric designed to capture and quantify temporal biomarker trajectories for the precise prediction of malignancy. In this large-scale, population-based cohort study conducted at the National Center for Respiratory Medicine in China, we analyzed real-world data from 12,459 participants—comprising 3472 patients with histologically confirmed lung cancer and 8987 with benign nodules—recruited between July 2013 and December 2023. We assayed serum concentrations of a comprehensive panel of six tumor biomarkers: carcinoembryonic antigen (CEA), neuron-specific enolase (NSE), carbohydrate antigens 125 and 153 (CA125, CA153), cytokeratin fragment 21–1 (CYFRA21–1), and squamous cell carcinoma antigen (SCC). Serial longitudinal data from a sub-cohort of 1847 patients who underwent at least three measurements over six months were utilized to construct the BVI using machine learning algorithms that optimized the weighting of individual marker velocities. Static biomarker analysis revealed that an optimized CEA threshold of 2.49 ng/mL significantly outperformed the conventional 5.0 ng/mL cutoff, increasing sensitivity from 42.2% to 69.8% (AUC 0.749). However, the dynamic BVI model demonstrated superior predictive performance (AUC 0.892) compared to static analysis alone (AUC 0.847). Patients stratified into the high-BVI category exhibited a 4.82-fold increased risk of malignancy (95% CI: 3.91–5.94, P 5.8 months. To maximize clinical utility, we proposed a two-stage diagnostic algorithm integrating initial static screening with BVI assessment for intermediate-risk patients. This integrated approach achieved 95.8% sensitivity and 82.4% specificity, reducing unnecessary examinations by 34.2% and saving approximately ¥20,498 per patient, with a number needed to screen of 28. BVI represents an innovative paradigm shift in lung nodule assessment. By leveraging temporal biomarker analysis, this dynamic framework offers a clinically applicable tool to revolutionize early lung cancer screening, optimize resource allocation, and improve patient outcomes through timely intervention.