Abstract Introduction Post-surgical recurrence affects up to 20% of patients with stage I non-small cell lung cancer (NSCLC), representing a primary cause of mortality and a significant clinical challenge. Accurate risk stratification to guide adjuvant therapy and surveillance is a critical unmet need. We hypothesized that a deep learning framework capable of integrating heterogeneous clinical, radiomic, and genomic data could provide a more accurate and personalized prediction of postoperative recurrence than unimodal approaches or traditional staging. This study aims to develop and validate MAGIC-Recur, a novel multi-modal, graph-based deep learning model for this purpose. Methods We retrospectively collected data from 631 patients who underwent curative surgical resection for stage I NSCLC. The dataset included clinical variables, preoperative CT images for radiomic feature extraction, and targeted sequencing data from a 56-gene panel. The MAGIC-Recur model was developed using parallel encoders for clinical and radiomic features, a Graph Transformer to model the interaction network among genes, and a cross-attention network for data fusion. The primary endpoint was recurrence-free survival (RFS). Model performance was evaluated using the Area Under the Receiver Operating Characteristic Curve (AUC-ROC), and its prognostic value was assessed using the Concordance Index (C-index) and Kaplan-Meier analysis. Results The integrated multi-modal MAGIC-Recur model demonstrated superior predictive performance compared to unimodal models. In the independent test set (n = 126), MAGIC-Recur achieved an AUC-ROC of 0.8924 (95% CI: 0.802-0.9829), outperforming models based on clinical (AUC 0.8241), genomic (AUC 0.8547), or radiomic (AUC 0.7781) data alone. The model demonstrated robust prognostic capability, achieving a C-index of 0.763 (95% CI: 0.704 - 0.816) for predicting RFS, which was comparable to the TNM staging system (p = 0.0949). It successfully stratified patients into high- and low-risk groups with a highly significant difference in survival outcomes (log-rank p 0.0001). The model’s graph-based genomic module provided key biological insights, identifying STK11 and MET as highly predictive of recurrence while revealing the prognostic value of the frequently mutated EGFR gene to be context-dependent within the network. Conclusions By effectively integrating clinical, radiomic, and graph-represented genomic data, the MAGIC-Recur model provides an accurate and interpretable prediction of recurrence risk in patients with resected stage I NSCLC. This multi-modal approach offers a more nuanced risk assessment than current standards and holds significant potential to enhance clinical decision-making by guiding personalized adjuvant therapies and surveillance strategies to improve patient outcomes. This abstract is funded by: This work was supported by National Natural Science Foundation of China (Nos. 92159302 to W Li)
Li et al. (Fri,) studied this question.