OBJECTIVE

To assess the effectiveness of a machine learning framework and nomogram in predicting progression-free survival (PFS) post-radical gastrectomy in patients with dMMR.

METHOD

Machine learning models and nomograms to forecast PFS in patients undergoing radical gastrectomy for nonmetastatic gastric cancer with dMMR. Independent risk factors were identified using Cox regression analysis to develop the nomogram. The performance of the models was assessed through C-index, time receiver operating characteristic (T-ROC) curves, calibration curves, and decision curve analysis (DCA) curves. Subsequently, patients were categorized into high-risk and low-risk groups based on the nomogram's risk scores.

RESULTS

Among the 582 patients studied, machine learning models exhibited higher c-index values than the nomogram. Random Survival Forests (RSF) demonstrated the highest c-index (0.968), followed by Extreme Gradient Boosting (XG boosting, 0.945), Decision Survival Tree (DST, 0.924), the nomogram (0.808), and 8th TNM staging (0.757). All models showed good calibration with low integrated Brier scores (< 0.1), although there was calibration drift over time, particularly in the traditional nomogram model. DCA showed an incremental net benefit from all machine learning models compared with conventional models currently used in practice. Age, positive lymph nodes, neural invasion, and Ki67 were identified as key factors and integrated into the prognostic nomogram.

CONCLUSION

Our research has demonstrated the effectiveness of the RSF algorithm in accurately predicting progression-free survival (PFS) in dMMR gastric cancer patients after gastrectomy. The nomogram created from this algorithm has proven to be a valuable tool in identifying high-risk patients, providing clinicians with important information for postoperative monitoring and personalized treatment strategies.