神经药理学报

神经药理学报 ›› 2026, Vol. 16 ›› Issue (3): 28-.DOI: 10.3969/j.issn.2095-1396.2026.03.004

• 研究论文 • 上一篇    下一篇

基于可解释机器学习的前列腺癌风险预测模型构建

曹丽，王菁菁，武丽媛，赵国斌   

  1. 1. 河北北方学院研究生学院，张家口，075000，中国 

    2. 河北北方学院附属第一医院泌尿外科，张家口，075000，中国

  • 出版日期:2026-06-26 发布日期:2026-08-04
  • 通讯作者: 赵国斌，主任医师，教授，硕士研究生导师；研究方向：泌尿科学
  • 作者简介:曹丽，硕士研究生；研究方向：前列腺癌
  • 基金资助:
    河北省自然科学基金资助项目（No.H2021405012）

Development and Validation of Machine Learning Models for Predicting Prostate Cancer Risk: A Multi-Cohort Study

CAO Li, WANG Jing-jing, WU Li-yuan, ZHAO Guo-bin   

  1. 1. Graduate School, Hebei North University, Zhangjiakou, 075000, China 

    2. Urology Surgery, the First Affiliated Hospital of Hebei North University, Zhangjiakou, 07500, China

  • Online:2026-06-26 Published:2026-08-04

PDF (PC)

3

摘要：

目的：通过使用临床医学数据，构建并验证多种机器学习预测模型，用于前列腺癌风险预测，旨在筛选 出最佳性能模型，并验证其在国内外独立数据集的泛化能力，为前列腺癌的临床决策提供稳健、可信的工具。方 法：选择国家人口健康科学数据中心的前列腺肿瘤患者临床数据，经高级特征筛选移除噪声数据后，按 7:3 划 分训练集和测试集，构建 5 种机器学习模型，包括逻辑回归（logistic regression，LR）、极端梯度提升（extreme gradient boosting，XGBoost）、随机森林（random forest，RF）、分类特征梯度提升（categorical boosting，CatBoost）、 K 近 邻（K-nearest neighbors，KNN）。 在 测 试 集 上 采 用 受 试 者 工 作 特 征 曲 线 下 面 积（area under the receiver operating characteristic curve，AUC）、准确率、召回率及 F1 分数综合评价模型性能，同时评估模型在国外患者独 立测试集的表现，最后利用 SHAP 对最佳模型进行特征重要性解析。结果：在五种模型中，XGBoost 模型在内部 测试集上表现最佳，其 AUC、准确率、召回率及 F1 分数分别为 0.894、0.852、0.825、0.834。SHAP 分析显示，总 PSA 水平、碱性磷酸酶、年龄、肌酸激酶同工酶、游离总前列腺特异性抗原（prostate specific antigen，PSA）比值 为最关键特征。同时，XGBoost 模型在独立测试集上同样有着优异的性能（AUC:0.810），证明了其强大的泛化能 力。结论：本研究构建了一个基于 XGBoost 的高性能、可解释的前列腺癌风险预测模型。该模型不仅在中国人 群中表现出色，同时在国际数据中也表现出良好的普适性，可以为临床医生提供精准的风险分层，从而优化诊疗 策略。

关键词: 前列腺癌, 机器学习, SHAP, XGBoost

Abstract:

Objective: To develop and validate machine learning models for predicting prostate cancer risk using clinical data, aiming to identify the optimal model and verify its generalizability across independent domestic and international cohorts, thereby providing a robust tool for clinical decision support. Methods: Clinical data of patients with prostate tumors were obtained from the National Population Health Science Data Center. Following advanced feature selection to eliminate noise, the data were randomly split into training and testing sets at a 7:3 ratio. Five machine learning models were constructed: logistic regression (LR), extreme gradient boosting (XGBoost), random forest (RF), categorical boosting (CatBoost), and K-nearest neighbors (KNN). Model performance was evaluated on the internal test set using the area under the receiver operating characteristic curve (AUC), accuracy, recall, and F1-score. The topperforming model was further validated on an independent external test set comprising international patient data. Finally, shapley additive exPlanations (SHAP) were employed to interpret the feature importance of the best model. Results: Among the five models, XGBoost demonstrated superior performance on the internal test set, with an AUC of 0.894, accuracy of 0.852, recall of 0.825, and an F1-score of 0.834. SHAP analysis revealed that total PSA, alkaline phosphatase, age, creatine kinase isoenzyme, and the free-to-total PSA ratio were the most critical predictive features. Notably, the XGBoost model maintained excellent performance on the independent external test set (AUC: 0.810), confirming its strong generalizability. Conclusion: This study developed a high-performance and interpretable XGBoost-based model for prostate cancer risk prediction. The model exhibited outstanding discriminative ability within the Chinese population and sustained good generalizability in an international cohort. It holds promise as a practical tool to assist clinicians in precise risk stratification and optimization of diagnosis and treatment strategies.

Key words: prostate cancer, machine learning, SHAP, XGBoost

中图分类号: