Radiomics-assisted unsupervised clustering and predictive modeling to identify optimal candidates for virtual...

Accurate localization of small pulmonary nodules during thoracoscopic lung resection remains challenging, and the selection of preoperative localization methods largely relies on surgeon experience. This study aimed to integrate radiomic features and clinical characteristics to develop a data-driven patient clustering approach and to explore the suitability of different preoperative localization methods across patient subgroups. Patients who underwent thoracoscopic wedge resection or segmentectomy with virtual navigation bronchoscopy (VNB) localization, computed tompgraphy (CT)-guided localization, or no localization were retrospectively included. Radiomic features were extracted from preoperative CT images and integrated with clinical and nodule characteristics. Unsupervised K-means clustering was applied. Within each cluster, perioperative outcomes were compared among localization methods. In addition, supervised learning models were constructed to predict cluster assignment and intraoperative blood loss. A total of 925 patients were included, and 3 stable clusters were identified. In Cluster 0, VNB localization was associated with lower intraoperative blood loss than CT-guided localization [wedge resection: 50 (20, 70) ml versus 70 (50, 100) ml, P = .004; segmentectomy: 50 (20, 100) ml versus 100 (50, 100) ml, P = .002]. In Cluster 2, CT-guided localization was associated with lower intraoperative blood loss than VNB localization [wedge resection: 50 (30, 90) ml versus 100 (50, 110) ml, P = .005; segmentectomy: 90 (50, 100) ml versus 105 (85, 170) ml, P = .011], along with a higher localization success rate. The supervised model demonstrated stable performance in predicting cluster assignment (macro-area under the curve = 0.863) and intraoperative blood loss (macro-area under the curve range: 0.813-0.850). VNB localization and CT-guided localization demonstrated distinct advantages across different clusters. These findings support a data-driven, individualized strategy for preoperative localization selection.