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).

Summary

A stacked ensemble integrating LDCT radiomics with cfDNA fragmentomics achieved AUC 0.950 internally and 0.966 externally, improving specificity at 95% sensitivity versus either modality alone. Performance gains were greatest for 10–20 mm solid nodules, and a companion model stratified invasiveness with AUC ≈0.88.

Key Findings

• The integrated model reached AUC 0.950 (internal) and 0.966 (external), outperforming imaging-only and cfDNA-only models.
• At 95% sensitivity, specificity improved to 0.60 vs 0.50 (imaging) and 0.33 (cfDNA), with marked gains for 10–20 mm and pure solid nodules.
• An invasiveness model stratified tumors with AUC ≈0.88, with scores increasing stepwise from AIS to invasive adenocarcinoma.

Clinical Implications

Integration of imaging and cfDNA analytics can be introduced into lung cancer screening programs to triage indeterminate nodules—improving specificity at a fixed high sensitivity, especially for 10–20 mm solid nodules.

Limitations

• Potential spectrum and referral biases and regional enrollment may limit generalizability.
• Prospective clinical utility, cost-effectiveness, and comparisons to standard risk models (e.g., Brock) were not reported.

Future Directions

Prospective impact and cost-effectiveness trials in screening cohorts, integration with clinical predictors, and evaluation across diverse populations and scanner protocols.