Development of a predictive model for pneumothorax after microwave ablation based on radiomics and clinical baseline data.

Aim: Lung cancer is a leading cause of cancer-related mortality globally, with a five-year survival rate lower than many other cancers. Surgery remains the most effective treatment; however, fewer than 50% of patients are eligible due to compromised pulmonary function or the presence of multiple lesions. Microwave ablation is an emerging, minimally invasive treatment that has shown promise in prolonging survival and preserving organ integrity with fewer side effects. Despite its safety profile, pneumothorax remains a common complication. Radiomics has gained traction for early diagnosis, prognosis prediction, and treatment assessment. This study aims to develop a predictive model for pneumothorax following MWA by integrating radiomic data.

Methods: Data from 111 lung cancer patients undergoing MWA were retrospectively analyzed. A clinical model was developed using binary logistic regression, while a radiomics model was constructed via LASSO regression with fivefold nested cross-validation. A comprehensive model was built by combining both feature sets using logistic regression. Model performance was evaluated using ROC curves, AUC values, DeLong's test, and calibration curves to assess the agreement between predicted and observed outcomes.

Results: The clinical model achieved an AUC of 0.8846 (95% CI: 0.8160-0.9533), the radiomics model had an AUC of 0.8353 (95% CI: 0.7453-0.9253), and the comprehensive model showed the highest AUC of 0.9262 (95% CI: 0.8712-0.9812). DeLong's test revealed that the comprehensive model outperformed both the clinical model (Z = -2.24, P = 0.025) and the radiomics model (Z = -2.57, P = 0.010).

Conclusion: Compared with the individual models, the predictive model developed by combining radiomic and clinical baseline data demonstrated superior diagnostic performance in predicting pneumothorax after microwave ablation. By incorporating additional multimodal data and clinical factors in the future, this model has the potential to serve as a more accurate predictive tool in clinical practice.