Risk-stratified classification of pulmonary nodule malignancy via a machine learning model integrating imaging and cell-free DNA: a model development and validation study (DECIPHER-NODL)

作者：Huiting Wang, Hairong Huang, Feng Li, Ying Deng, Changyong Wang, Wei Wei, Song Wang, Dongqin Zhu, Hao Xu, Hua Bao, Zheng Li, Wenjun Ye, Yuan Zhang, Caichen Li, Bo Cheng, Xiwen Liu, Liping Liu, Zheng Li, Jing Yang, Wei Chen, Peng He, Fufeng Wang, Wen Zhong, Weisheng Guo, Yang Shao, Yi Shen, Jianxing He, Wenhua Liang · 发表于：The Lancet Regional Health - Western Pacific · 年份：2025 · DOI：10.1016/j.lanwpc.2025.101730 · 被引用次数：2 · 研究领域：Lung Cancer Diagnosis and Treatment、Radiomics and Machine Learning in Medical Imaging、Cancer Genomics and Diagnostics

Background: Accurate risk stratification of pulmonary nodules is critical for early lung cancer detection. This study aimed to improve malignancy classification and invasiveness prediction using machine learning models integrating low-dose computed tomography (LDCT) radiomics and plasma cell-free DNA (cfDNA) fragmentomics. Methods: This multicenter study enrolled 1356 participants across discovery (n = 1147) and external validation (n = 209) cohorts. A deep learning-based imaging model processed LDCT scans for automated lung nodule detection and malignancy classification. A parallel cfDNA model analyzed four whole-genome fragmentation features: copy number variation, fragment size ratio, fragment-based methylation, and mutation context and signature. The two models were integrated via a stacked ensemble algorithm. An invasion prediction model evaluated tumor aggressiveness. Findings: The integrated imaging-cfDNA model outperformed individual models, with an AUC of 0.950 (95% CI: 0.926-0.975) in the internal test set and 0.966 (95% CI: 0.940-0.991) in the external validation. The combined model's specificity increased to 0.60 (95% CI: 0.49-0.71) while maintaining 95% sensitivity, compared to specificities of 0.50 (95% CI: 0.41-0.59) and 0.33 (95% CI: 0.23-0.44) at equivalent sensitivity levels for the imaging and cfDNA models, respectively. The combined model consistently outperformed the other two models across nodule characteristics, with particular improvement for 10-20 mm ...