Development and External Validation of ECMP-Net: A Deep Learning Framework for Predicting Brain Metastasis in Esophageal Cancer
Hongyu Yu, Wuxue Yu, Lihua RenObjective
Brain metastasis in esophageal cancer is uncommon but clinically severe, and externally validated prediction tools remain limited. This study aimed to develop and externally validate ECMP-Net, a multilayer perceptron model based on structured clinical predictors, for estimating brain metastasis risk in patients with esophageal cancer.
Methods
This retrospective model-development and external-validation study used the Surveillance, Epidemiology, and End Results (SEER) database from 2010 to 2020 and an external validation cohort from the First Affiliated Hospital of Jinzhou Medical University, China, from 2022 to 2024. A total of 19,502 eligible SEER patients were randomly divided into training and internal test sets. The external validation cohort included 274 patients, including 23 with brain metastasis. ECMP-Net used eight structured clinical predictors and was compared with decision tree, k-nearest neighbors, support vector machine, random forest, logistic regression, XGBoost, LightGBM, and CatBoost. Performance was assessed using AUC, AUPRC, Brier score, calibration, bootstrap confidence intervals, and threshold-dependent metrics.
Results
In the internal test set, ECMP-Net achieved an AUC of 0.813, AUPRC of 0.345, and Brier score of 0.064. In the external validation cohort, ECMP-Net achieved an AUC of 0.714, AUPRC of 0.224, and Brier score of 0.077. ECMP-Net showed the highest internal AUC and AUPRC and the highest external AUC, whereas logistic regression showed the highest external AUPRC and the lowest external Brier score. At the fixed training-derived threshold, external sensitivity, specificity, PPV, NPV, and positive-class F1-score were 0.609, 0.753, 0.184, 0.955, and 0.283, respectively.
Conclusion
ECMP-Net showed acceptable discrimination and competitive overall performance for brain metastasis risk prediction in esophageal cancer. Given the limited number of external events, larger multicenter prospective validation is required before clinical implementation.