Development and evaluation of multi-strategy fusion machine learning models for predicting postoperative recurrence in non-small cell lung cancer

Abstract Background We developed multiple machine learning models using various fusion strategies to predict the risk of recurrence within three years post-surgery in non-small cell lung cancer (NSCLC) patients. Our aim was to enhance the accuracy of prognostic predictions and assess the differences and effectiveness of these fusion strategies. Methods We retrospectively included 306 patients with histologically confirmed NSCLC who underwent curative surgical resection. Preoperative clinical variables and CT-based radiomic features were extracted for analysis. Radiomic features were harmonized using the ComBat method to mitigate scanner-related variability. Feature selection was performed within a nested five-fold cross-validation framework, and features consistently selected in at least three inner folds were retained. Six predictive models were developed, including a clinical model, a radiomics model, an early fusion model, two intermediate fusion models (feature-level fusion and radiomics score (Radscore)-level fusion), and a late fusion model. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, and F1-score. Model discrimination was compared using DeLong’s test, and SHAP analysis was applied to enhance model interpretability. Results Among all developed models, fusion-based approaches consistently outperformed unimodal models. The Radscore-level fusion model demonstrated the best discriminative performance, achieving the highest cross-validated AUC of 0.864, followed closely by the feature-level fusion model (AUC = 0.860). Both intermediate fusion models significantly outperformed the clinical model and showed superior discrimination compared with the radiomics model. The early fusion and late fusion models yielded comparable but relatively lower performance. DeLong test results confirmed that the Radscore-level fusion model significantly outperformed the clinical, radiomics, early fusion, and late fusion models. SHAP analysis further illustrated the relative importance and directional effects of key clinical and radiomic features contributing to model predictions. Conclusions Multiple fusion strategies are feasible and effective for predicting postoperative recurrence in patients with NSCLC. Among them, the feature-level fusion approach maintains favorable predictive performance while offering improved interpretability, supporting its potential utility for postoperative recurrence risk stratification.