Explainable machine learning for the prognostication of salivary duct carcinoma: Development and deployment of a web-based prediction tool.

BACKGROUND

Salivary duct carcinoma (SDC) is a rare but aggressive malignancy often associated with lymph node metastasis and poor prognosis. Therefore, although accurate prognostication is crucial, traditional models are often inadequate because of linear assumptions and limited interpretability. In contrast, machine learning (ML) offers a flexible, interpretable framework for improving survival prediction and supporting individualized care planning.

METHOD: Overall, 552 patients with SDC (2004-2021) were identified from the Surveillance, Epidemiology, and End Results database and stratified by cancer-specific survival (CSS) and overall survival (OS) status, before being split into the training and testing sets (7:3). Three prognostic models were developed: Cox proportional hazards, random survival forest (RSF), and DeepSurv. Model performance was evaluated using the concordance index (C-index), integrated Brier score, time-dependent area under the curve (AUC), calibration curves, and decision curve analysis. Shapley additive explanations (SHAP) values were applied to enhance model interpretability and quantify the contribution of individual features to risk prediction.

RESULT: All three models demonstrated favorable predictive performance, with the RSF model showing the best discrimination and calibration (C-index: 0.785 in training and 0.768 in testing). For CSS prediction, the 1-, 3-, and 5-year AUCs in the testing set were 0.781, 0.810, and 0.818, respectively. SHAP analysis identified positive lymph node ratio, TNM stage, and excision surgery as key prognostic predictors. The RSF model was selected for deployment as an interactive web-based tool.

CONCLUSION

This study established an interpretable ML-based model that reliably predicts CSS and OS in patients with SDC. Its successful deployment as a web-based tool underscores its potential to enhance personalized prognostic assessment and support evidence-based clinical management.