Objectives: To develop and externally validate an interpretable fusion model combining multi-time-point CT radiomics with clinical-semantic features to predict invasiveness of pulmonary ground-glass nodules and support three-tier risk stratification.
Materials and methods: In this multicentre retrospective study, patients with pulmonary ground-glass nodules that were resected or managed via CT surveillance with stability (≥ 3 years) were included. Thin-section CT scans at baseline (T0) and follow-up (T1) were used to derive radiomic features at each time point and delta-radiomic features. Four unimodal models (T0 radiomics, T1 radiomics, delta-radiomics, and clinical-semantic) were trained using centre-grouped cross-validation and probability calibration, then fused via stacked logistic regression. Low- and high-risk groups defined by two training-derived and locked probability thresholds were evaluated in an external cohort against clinical and guideline-based models.
Results: The training and external validation cohorts included 358 and 46 patients, respectively. The fusion model achieved an external area under the receiver operating characteristic curve of 0.985 (95% CI: 0.955-1.000) with good calibration. Using training-derived and locked thresholds (0.50 and 0.65), 28.3% of patients were classified as low risk (NPV 100%; 95% CI: 75.3-100.0) and 69.6% as high risk (PPV 93.8%; 95% CI: 79.2-99.2; sensitivity 96.8%; 95% CI: 83.3-99.9). The model reduced false-positive high-risk classifications from 33.3 to 13.3 per 100 non-invasive lesions and showed higher net benefit than comparator models.
Conclusions: An interpretable fusion model enables robust three-tier risk stratification of pulmonary ground-glass nodules and may reduce overdiagnosis and overtreatment in low-dose CT screening programmes.
Critical relevance statement: A calibrated, interpretable fusion model for pulmonary ground-glass nodules enables accurate three-tier risk stratification, reducing false-positive high-risk classifications and supporting safer de-escalation of surveillance in low-risk patients.
Key points: An interpretable fusion model combining multi-time-point CT radiomics with clinical-semantic features predicts invasiveness of pulmonary ground-glass nodules. The fusion model achieved excellent discrimination (external AUC 0.985) and good calibration, outperforming unimodal radiomics and guideline-based risk models. A three-tier risk stratification with calibrated thresholds reduces false-positive high-risk classifications and supports safe de-escalation of surveillance in low-risk patients.
Keywords: Ground-glass nodule; Lung neoplasms; Machine learning; Radiomics; Tomography (X-ray computed).
© 2026. The Author(s).