Background
Low-dose computed tomography (LDCT) is the only recommended noninvasive modality for lung cancer screening and has been shown to reduce lung cancer mortality. However, it is accompanied by a high prevalence of pulmonary nodules, particularly in East Asian populations, where over 90% of detected nodules are ultimately benign. Existing guideline-based strategies and conventional malignancy prediction models rely heavily on nodule size and a limited number of manually assessed features, resulting in suboptimal discrimination and high false-positive rates. Although artificial intelligence (AI) methods have improved imaging-based classification, their performance often declines in real-world screening settings and they frequently ignore broader epidemiological and biological determinants of malignancy. There is therefore a critical need for validated, reproducible, and multimodal risk stratification tools tailored to large-scale screening populations.
 
Objectives
To develop and validate single-modality and multi-modality models for pulmonary nodule malignancy classification based on a single LDCT scan, and to evaluate their performance against established risk models and real-world radiologist assessments in large prospective Chinese cohorts.
 
Methods
Model development was conducted in a hospital-based case–control dataset including 935 surgically confirmed malignant nodules and 935 matched benign nodules, with external validation in two independent prospective cohorts from the Lung Imaging Genomics Initiative (LIGI) comprising 28,185 screening participants and 12,007 health examination participants. Pulmonary nodules were automatically detected and segmented, and quantitative imaging features were extracted using an AI-assisted pipeline. A single-modality nodule malignancy score was constructed using six machine-learning algorithms. A multi-modality score was further developed by integrating the nodule score with AI-derived malignancy probability, epidemiological lung cancer risk score (LCRS), polygenic risk score (PRS), and lobar radiomic features, using ridge regression. Model performance was evaluated using AUC, sensitivity, specificity, positive and negative predictive values, and the net reclassification index (NRI), and compared against community radiologists’ LDCT assessments.
 
Results
The single-modality nodule malignancy score demonstrated robust and consistent discrimination in external validation cohorts (AUC 0.905 and 0.870), outperforming mainstream models including Mayo, VA, and PKUPH. Participants in the highest decile of the score showed a lung cancer incidence of 9.29% and accounted for over 70% of malignant cases within two years. The multi-modality score further improved discrimination (AUC up to 0.955), with significant risk reclassification gains (NRI: 3.72–14.71%) over the single-modality model, particularly among never-smokers. Compared with manual radiologist assessment, the multi-modality score achieved a true-positive rate of 84.19%, reducing false negatives by approximately 30% while maintaining acceptable false-positive rates. Performance gains were consistent across subgroups defined by smoking status, nodule size, density, and follow-up duration.
 
Conclusions
By integrating nodule features with multidimensional risk factors, we developed robust single- and multi-modality models that substantially improve pulmonary nodule malignancy classification from a single LDCT scan. These models demonstrate strong performance in large real-world screening cohorts and offer a practical, noninvasive tool to enhance risk stratification, reduce missed cancers, and support precision lung cancer screening and management.

Flowchart of participant selection, data preprocessing, model construction, and validation