Wang Eric, Shuryak Igor, Brenner David J
Center for Radiological Research, Department of Radiation Oncology, Columbia University Irving Medical Center, New York, United States of America.
PLoS One. 2025 Jul 21;20(7):e0328082. doi: 10.1371/journal.pone.0328082. eCollection 2025.
This study employed state-of-the-art machine learning to evaluate the mortality effects of alpha-emitting radionuclides (241Am, 249Cf, 252Cf, 238Pu, 239Pu, 224Ra, 226Ra, 228Th) on 2,576 dogs, factoring in radioactivity levels, composition, administration method (injection or inhalation), and age at exposure. There were 972 cancer deaths, 599 non-cancer deaths, 789 deaths from many diseases (involving several diagnoses, including both cancer and non-cancer pathologies), and 216 deaths with uncertain causes. A Random Survival Forest model for overall mortality achieved concordance scores of 0.763 and 0.745 on training and testing data subsets, respectively. A model variant with competing risks was used to investigate mortality trends over time for different disease categories. It achieved concordances of 0.814 for cancer, 0.652 for non-cancer, and 0.778 for many diseases on training data, and 0.817 for cancer, 0.651 for non-cancer, and 0.780 for many diseases on testing data. All radionuclides exhibited radiation responses for cancer, with 226Ra and 239Pu showing the strongest effects. Some responses were non-linear, with indications of saturation or downturn at high treatment quantities. For non-cancer diseases, radiation responses were generally weaker and more variable. For the many diseases endpoint, 238Pu and 239Pu demonstrated the strongest response patterns, with 239Pu exhibiting greater lethality via inhalation compared to injection.. Using a Causal Forest model, which is designed to detect causal relationships rather than just associations, we investigated the causal impact of radioactivity on dog mortality, accounting for other variables. We found a significant (p < 2 × 10-16) negative average causal effect of -1,375 days per log10 radioactivity unit on survival time. This study improves current knowledge of cancer and non-cancer mortality patterns from densely-ionizing radiation in mammals by using machine learning to analyze combined historical data on dogs exposed to different radionuclides, modeling multiple variables, nonlinear dependencies, and causal relationships.
本研究采用了最先进的机器学习技术，以评估发射α粒子的放射性核素（241Am、249Cf、252Cf、238Pu、239Pu、224Ra、226Ra、228Th）对2576只狗的致死效应，同时考虑了放射性水平、成分、给药方式（注射或吸入）以及暴露时的年龄。其中有972例癌症死亡、599例非癌症死亡、789例多种疾病导致的死亡（涉及多种诊断，包括癌症和非癌症病理）以及216例死因不明的死亡。一个用于总体死亡率的随机生存森林模型在训练和测试数据子集上分别获得了0.763和0.745的一致性得分。一个具有竞争风险的模型变体用于研究不同疾病类别的死亡率随时间的变化趋势。在训练数据上，它在癌症方面的一致性为0.814，非癌症方面为0.652，多种疾病方面为0.778；在测试数据上，癌症方面为0.817，非癌症方面为0.651，多种疾病方面为0.780。所有放射性核素对癌症均表现出辐射反应，其中226Ra和239Pu的效应最强。一些反应是非线性的，在高治疗剂量时显示出饱和或下降的迹象。对于非癌症疾病，辐射反应通常较弱且更具变异性。对于多种疾病终点，238Pu和239Pu表现出最强的反应模式，与注射相比，239Pu通过吸入表现出更高的致死率。使用旨在检测因果关系而非仅仅关联的因果森林模型，我们在考虑其他变量的情况下，研究了放射性对狗死亡率的因果影响。我们发现，每log10放射性单位对生存时间的平均因果效应为-1375天，具有显著意义（p < 2×10-16）。本研究通过使用机器学习分析暴露于不同放射性核素的狗的综合历史数据、对多个变量、非线性依赖性和因果关系进行建模，提高了我们对哺乳动物中密集电离辐射导致的癌症和非癌症死亡率模式的现有认识。
