可解释机器学习识别影响前列腺癌生存结局的卫生系统杠杆
该研究运用可解释机器学习(CatBoost与SHAP)分析全球185个国家前列腺癌死亡发病比(MIR)与11项卫生系统指标的关系。模型预测性能良好(R²=0.80),发现人均GDP、全民健康覆盖服务覆盖指数和放疗基础设施与较低MIR最强相关,而单纯增加卫生总支出关联较弱。
要点
- 人均GDP、全民健康覆盖指数和放疗基础设施与较低前列腺癌MIR最强相关。
- 单纯提高卫生总支出与改善前列腺癌结局的关联较弱。
- 国家层面SHAP分解显示特征贡献异质,需结合情境确定优先事项。
核心结论
- 放疗基础设施、全民健康覆盖和国家财富是与较低前列腺癌死亡发病比最一致的卫生系统特征。
- 针对性投资治疗能力和服务覆盖可能比单纯增加卫生总支出对前列腺癌结局更有意义。
研究设计与样本
- 应用可解释机器学习(CatBoost与SHAP)分析GLOBOCAN 2022中全部185个国家的前列腺癌MIR估计。
- 将年龄标化MIR与来自WHO全球卫生观察站、世界银行、联合国机构和放疗中心目录的11项卫生系统指标关联。
- 采用重复留一国家交叉验证和基于bootstrap的不确定性估计评估模型性能,CatBoost原生处理缺失预测值。
主要发现
- 模型对前列腺癌MIR预测性能强:R²=0.80,RMSE=0.078,MAE=0.060,Pearson相关系数=0.89。
- 与较低MIR最强相关的卫生系统特征为人均GDP、全民健康覆盖服务覆盖指数和放疗基础设施。
- 国家层面SHAP分解显示特征贡献异质,可识别情境特异性优先事项。
- 更大放疗能力和更全面全民健康覆盖与较低MIR一致相关,而较高卫生总支出单独与结局改善关联较弱。
临床怎么用
- 提示卫生系统决策者可优先投资放疗基础设施和全民健康覆盖服务覆盖以改善前列腺癌结局。
- 单纯增加卫生总支出可能不足以缩小前列腺癌生存差异。
- 该框架可为未来卫生系统研究、前瞻性评估和全球前列腺癌控制提供参考。
局限性
- 研究为生态学设计且属假设生成性质。
- 结论需未来前瞻性评估验证。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
BACKGROUND: Advances in the care of patients with prostate cancer across the disease spectrum have resulted in substantial survival gains over recent decades. However, global disparities in prostate cancer survival, stage at diagnosis, and access to timely diagnosis and treatment remain among the greatest across all malignancies. We aimed to quantify country-level health system correlates of the prostate cancer mortality-to-incidence ratio (MIR). METHODS: We applied interpretable machine learning using CatBoost with SHAP (SHapley Additive exPlanations) to all 185 countries with prostate cancer MIR estimates in GLOBOCAN 2022. National age-standardized MIR estimates were linked to 11 health system indicators from the World Health Organization Global Health Observatory, World Bank World Development Indicators, United Nations agencies, and the Directory of Radiotherapy Centres. Model performance was evaluated using repeated leave-one-country-out cross-validation with bootstrap-based uncertainty estimation. CatBoost handled missing predictor values natively, allowing inclusion of all countries without imputation. RESULTS: The model demonstrated strong predictive performance for prostate cancer MIR (R² = 0.80, RMSE = 0.078, MAE = 0.060, Pearson correlation = 0.89). The health system features most strongly associated with lower MIR were gross domestic product per capita, the universal health coverage (UHC) service coverage index, and radiotherapy infrastructure. Country-level SHAP decompositions demonstrated heterogeneous feature contributions across settings, identifying context-specific priorities. Greater radiotherapy capacity and more comprehensive UHC were consistently associated with lower MIR, whereas higher aggregate health expenditure alone showed weaker associations with improved outcomes. CONCLUSIONS: Interpretable machine learning identified radiotherapy infrastructure, universal health coverage, and national wealth as the health system features most consistently associated with lower prostate cancer mortality-to-incidence ratios across countries. These findings suggest that targeted investment in treatment capacity and service coverage may have greater relevance to prostate cancer outcomes than increases in aggregate health spending alone. Although ecological and hypothesis-generating, this framework may inform future health systems research, prospective evaluation, and global prostate cancer control.
引用
- Patel MS, Janopaul-Naylor J, Ting FIL, et al. Interpretable machine learning identifies health system levers for survival outcomes of patients with prostate cancer. Prostate Cancer Prostatic Dis. 2026. PMID: 42827104. doi:10.1038/s41391-026-01156-x
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。