What question did this study set out to answer?

July 20, 2026Open Access

Radiomic model with SHAP-based interpretability for predicting invasiveness of pure ground-glass nodules: a retrospective study based on high-resolution computed tomography (HRCT) volumetric datasets

HSHui ShengTianjin Academy of Agricultural Sciences RWRui Wang GZGuowei Zhang

Key Points

This research aims to develop an interpretable model to predict the invasiveness of pure ground-glass nodules using radiomic and clinical data.
Retrospective analysis of 235 surgically resected pGGNs classified by invasiveness according to WHO guidelines.
Developed clinical, radiomic, and combined prediction models employing LASSO regression for feature selection.
Model performance assessed using AUC, calibration curves, decision curve analysis, and SHAP for feature importance.
Combined model achieved AUC of 0.923 in training and 0.832 in testing, outperforming clinical and radiomic models.
Decision curve analysis confirmed the clinical utility of the combined model.
SHAP analysis identified log_sigma_2_0mm_3D_firstorder_Range as the most influential predictive feature.

Abstract

Abstract Background Radiomics holds promise for lung cancer diagnosis. This study developed an interpretable radiomics–clinical model to predict the invasiveness of pure ground-glass nodules (pGGNs) on high-resolution computed tomography (HRCT). To address the model’s “black box” nature, we applied the SHapley Additive exPlanations (SHAP) framework. Methods We retrospectively analyzed 235 surgically resected, histopathologically confirmed pGGNs, classified as non-invasive (AAH/AIS/MIA) or invasive (IAC) according to the 2015 WHO classification of lung tumors. We developed three prediction models: clinical, radiomic, and combined. Feature selection for the radiomic and combined models employed LASSO regression. Model performance was assessed using AUC, calibration curves, and decision curve analysis (DCA). Additionally, SHAP was used to quantify feature importance and to generate individualized explanations. Results Two clinico-radiological features (mean CT value, VolumePercent₋₃₀₀) and eight radiomic features were retained. The combined model yielded AUCs of 0. 923 (training) and 0. 832 (testing), outperforming the clinical model (0. 799/0. 733) and the radiomic model (0. 917/0. 827). Decision curve analysis (DCA) confirmed the superior clinical utility of the combined model. SHAP analysis ranked logₛigma₂₀mm₃DfirstorderRange as the single most important predictive feature. Conclusions The SHAP-augmented radiomics–clinical model offers an accurate and interpretable preoperative assessment of pGGN invasiveness. This tool can help clinicians choose the optimal surgical strategy and support individualized decision-making.

Ask AI

Helpful

Bookmark

View Full Paper