Integrating multimodal features to predict the malignancy of pulmonary ground-glass nodules: a multicenter prospective model development and validation study.

Summary

Among 571 GGNs (501 participants) across seven centers, the clinic–biomarker–deep radiomic (CB-DR) model achieved an AUC of 0.90 (95% CI 0.81–0.97) in the test set with accuracy 0.89, sensitivity 0.90, and specificity 0.82. Model-guided decisions would have reduced overtreatment in 82.4% of benign GGNs and enabled timely intervention in 90% of malignant GGNs.

Key Findings

• CB-DR model achieved AUC 0.90 with 0.89 accuracy, 0.90 sensitivity, and 0.82 specificity in external testing.
• Integrating biomarkers with deep radiomics and clinical features outperformed unimodal models.
• Decision analysis suggests substantial reduction in overtreatment of benign GGNs (82.4%) and improved capture of malignant GGNs (90%).

Clinical Implications

A calibrated risk tool could triage GGNs, reduce unnecessary resections and follow-up imaging, and prioritize timely intervention for high-risk lesions.

Limitations

• Geographic concentration in one country may limit generalizability
• Biomarker panels and imaging protocols may require standardization across sites

Future Directions

Head-to-head evaluation against established clinical risk models, prospective impact studies on management decisions and outcomes, and multi-national validation.