ObjectiveTo develop a multimodal recurrence prediction model for stage I NSCLC by integrating radiomics, genomics, and clinical data through hierarchical feature engineering and cross-omics interaction algorithms.MethodsThis study pioneered a hierarchical multimodal integration framework, synergizing radiomics (2 265 nodule features), genomics (116 whole-exome sequencing profiles), and clinical data from 323 stage I NSCLC patients. A two-stage feature engineering pipeline (correlation analysis + Random Forest selection) optimized discriminative features, and a cross-omics interaction algorithm dynamically quantified spatial associations between imaging phenotypes and genomic alterations. ResultsSingle-modality model performance: The radiomics model (XGBoost, AUC=0.896±0.088) demonstrated the best predictive efficacy among single-modality models, significantly outperforming the genomics-based model (Random Forest, AUC=0.644±0.196) and the clinical model (Random Forest, AUC=0.742±0.160). Dual-modality model performance: The radiomics-genomics integrated model achieved the best performance (XGBoost, AUC=0.942±0.077) among dual-modality models, superior to the radiomics-clinical model (XGBoost, AUC=0.929±0.086) and the genomics-clinical model (Random Forest, AUC=0.733±0.164).Trimodal integration performance: The trimodal model integrating radiomics, genomics, and clinical data achieved the peak performance (XGBoost, AUC=0.971±0.082), which was significantly better than all single-modality and dual-modality benchmarks.ConclusionsThis work establishes a methodological framework for predicting postoperative recurrence in stage I NSCLC through multimodal feature integration guided by domain-specific machine learning. Its superior performance highlights the advantage of multimodal integration, offering actionable insights for personalized surveillance strategies.