Journal: Nature communications
This study describes the development and validation of a multimodal prediction model, Recurrence Stratification and Assessment (RSA), to estimate early postoperative recurrence risk in patients with locally advanced gastric cancer after curative-intent surgery.
Key points:
- Problem: Early postoperative recurrence remains common in locally advanced gastric cancer, and standard clinicopathologic staging does not adequately capture biological heterogeneity in recurrence risk.
- Model design:
- RSA integrates deep learning–extracted histopathologic features from routine H&E-stained slides with conventional clinical variables.
- The model is designed to be clinically interpretable, using Shapley Additive Explanations (SHAP) to highlight which histologic patterns most influence recurrence risk.
- Cohorts and performance:
- Developed in a large, retrospective, multicenter cohort (n=1,763).
- Validated across two internal validation cohorts, two independent external cohorts from different geographic regions, and explored in a post‑hoc analysis of a prospective clinical trial population (NCT01516944).
- Demonstrated consistently strong discrimination for early recurrence with AUCs between 0.843 and 0.887 across datasets, suggesting robustness and generalizability.
- Biological insights:
- Transcriptomic sequencing and immune profiling were performed on available tumor samples.
- Tumors classified as low-risk by RSA showed immune-enriched microenvironments and higher expression of immune checkpoint genes, indicating more active antitumor immune responses compared with higher-risk tumors.
- This implies that differential immune biology may underlie the model’s risk stratification and the observed recurrence patterns.
- Clinical implications:
- RSA offers a high-performing, explainable digital pathology tool that could refine postoperative risk stratification beyond standard staging.
- Potential applications include tailoring surveillance intensity and serving as a biologically informed framework to investigate whether immune checkpoint inhibitors or other immunotherapies are differentially beneficial across risk groups.