A prognostic model for highly aggressive prostate cancer using interpretable machine learning techniques.

BACKGROUND

Extremely aggressive prostate cancer, including subtypes like small cell carcinoma and neuroendocrine carcinoma, is associated with poor prognosis and limited treatment options. This study sought to create a robust, interpretable machine learning-based model that predicts 1-, 3-, and 5-year survival in patients with extremely aggressive prostate cancer. Additionally, we sought to pinpoint key prognostic factors and their clinical implications through an innovative method.

MATERIALS AND METHODS

This study retrospectively analyzed data from 1,620 patients with extremely aggressive prostate cancer in the SEER database (2000-2020). Feature selection was performed using the Boruta algorithm, and survival predictions were made using nine machine learning algorithms, including XGBoost, logistic regression (LR), support vector machine (SVM), random forest (RF), k-nearest neighbor (KNN), decision tree (DT), elastic network (Enet), multilayer perceptron (MLP) and lightGBM. Model performance was evaluated using metrics such as AUC, accuracy (F1 score), confusion matrix, and decision curve analysis. Additionally, Shapley Additive Explanations (SHAP) were applied to interpret feature importance within the model, revealing the clinical factors that influence survival predictions.

RESULTS

Among the nine models, the lightGBM model exhibited the best performance, with an AUC and F1 score of (0.8, 0.809) for 1-year survival prediction, (0.809, 0.751) for 3-year survival prediction, and (0.773, 0.611) for 5-year survival prediction. SHAP analysis revealed that M stage was the most important feature for predicting 1- and 3-year survival, while PSA level had the greatest impact on 5-year survival predictions. The model demonstrated good clinical utility and predictive accuracy through decision curve analysis and confusion matrix.

CONCLUSION

The lightGBM model has good predictive power for survival in patients with extremely aggressive prostate cancer. By identifying key clinical factors and providing actionable predictions, the model has the potential to enhance prognostic accuracy and improve patient outcomes.