Machine learning–optimized multi-biomarker panel for discriminating lung cancer from LDCT-positive benign pulmonary nodules.

Abstract

10539 Background: Accurate discrimination between lung cancer and benign pulmonary nodules remains a critical challenge in lung cancer screening, particularly among individuals with positive findings on low-dose computed tomography (LDCT). Blood-based biomarkers with proven biological relevance may improve risk stratification when integrated with advanced analytical approaches. Serum carcinoembryonic antigen (CEA), serum amyloid A (SAA), and osteopontin (OPN) are biomarkers reflecting tumor burden, host inflammatory response, and tumor microenvironmental activity, respectively. The discriminatory performance of these three biomarkers between healthy individuals and lung cancer patients has previously been validated in a cohort exceeding 2,000 subjects (ASCO 2025, Abstract #489322). Building on this foundation, the present study evaluates whether algorithmic integration of these biomarkers can improve discrimination of lung cancer within the clinically challenging population of LDCT-positive pulmonary nodules. Methods: Serum CEA, SAA, and OPN levels were quantified by ELISA in patients with histologically confirmed lung cancer and individuals with LDCT-positive benign pulmonary nodules. To enhance discrimination within this high-risk cohort, machine learning–based modeling was applied. Multiple analytical methods were evaluated, and a Random Forest classifier was selected to capture inter-marker correlations and non-linear biological relationships. The dataset was divided into independent training and validation cohorts. Diagnostic performance was assessed using sensitivity, specificity, and area under the receiver operating characteristic curve (AUC). Results: In the LDCT-positive pulmonary nodule population, the machine learning–optimized three-biomarker panel demonstrated strong discriminatory performance. In the validation cohort, the model achieved an AUC of 0.879, with a sensitivity of 88.8% and a specificity of 83.3% for differentiating lung cancer from benign nodules. The algorithm consistently outperformed individual biomarkers and linear models, including in early-stage lung cancer. Conclusions: Application of machine learning to a biologically complementary three-protein biomarker panel significantly enhances discrimination of lung cancer within LDCT-positive high-risk pulmonary nodule populations. These findings support the clinical utility of this non-invasive approach as an adjunct to LDCT, addressing a key diagnostic gap in lung cancer screening and management.