Related Experiment Video
Updated: Sep 7, 2026
Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
Ensemble machine learning classifiers based on computed tomography radiomics for predicting spread through air spaces
Hongliang Qi1, Wanyin Qi2, Sanhong Zhang3
1Department of Clinical Engineering, Nanfang Hospital, Southern Medical University, Guangzhou, China.
Background:
Spread through air spaces (STAS) in lung adenocarcinoma (LUAD) is associated with adverse outcomes and may have implications for surgical planning. We aimed to develop and externally validate an ensemble machine learning model integrating preoperative computed tomography (CT) radiomics and clinicoradiological (CR) features for the prediction of STAS.
Methods:
This multicenter retrospective study included 1,206 patients with stage I LUAD from three centers, of whom 384 had STAS-positive tumors. Patients from two centers were divided into a training set (n=675) and an internal test set (n=290), and patients from the third center formed an external validation cohort (n=241). A radiomics score (Rad-score) was constructed using least absolute shrinkage and selection operator (LASSO) regression. Six tree-based algorithms and three ensemble strategies were evaluated using the Rad-score and CR features. The fixed Rad-score formula and threshold derived from the training set were transferred to a public radiogenomic cohort with matched CT images, whole-slide images (WSIs), and RNA sequencing data.
Results:
The stacking model achieved the best performance, with area under the receiver operating characteristic curve (AUC) values of 0.932 [95% confidence interval (CI): 0.904-0.960] in the internal test set and 0.883 (95% CI: 0.843-0.923) in the external validation cohort. SHapley Additive exPlanations (SHAP) analysis identified the Rad-score, CT density, and nodule size as the most influential predictors of model-predicted STAS risk. In the radiogenomic cohort, the transferred Rad-score classification was concordant with pathologic STAS status in 19 of 24 cases (Fisher's exact test, P=0.0056). Radiomics-predicted high-risk tumors showed transcriptomic alterations related to cell junction assembly, cell adhesion molecule (CAM) pathways, MYC targets, and glycolysis.
Conclusions:
A stacking ensemble model combining CT radiomics and CR features enabled robust, noninvasive preoperative prediction of STAS in stage I LUAD. Exploratory radiogenomic analysis suggested that the imaging-derived high-risk signature was associated with decreased cellular adhesion and metabolic reprogramming.