Cox Louis Anthony, Lewis R Jeffrey, Rege Saumitra V, Singh Shubham
Cox Associates, Entanglement, and University of Colorado. 503 N. Franklin Street, Denver, Colorado, 80218, USA.
Kelly Services, Epidemiology Contractor (retired ExxonMobil Biomedical Sciences, Inc.), Lavallette, New Jersey, USA.
Glob Epidemiol. 2024 Dec 25;9:100179. doi: 10.1016/j.gloepi.2024.100179. eCollection 2025 Jun.
AI-assisted data analysis can help risk analysts better understand exposure-response relationships by making it relatively easy to apply advanced statistical and machine learning methods, check their assumptions, and interpret their results. This paper demonstrates the potential of large language models (LLMs), such as ChatGPT, to facilitate statistical analyses, including survival data analyses, for health risk assessments. Through AI-guided analyses using relatively recent and advanced methods such as Individual Conditional Expectation (ICE) plots using Random Survival Forests and Heterogeneous Treatment Effects (HTEs) estimated using Causal Survival Forests, population-level exposure-response functions can be disaggregated into individual-level exposure-response functions. These reveal the extent of heterogeneity in risks across individuals for different levels of exposure, holding other variables fixed. By applying these methods to an illustrative dataset on blood lead levels (BLL) and mortality risk among never-smoker men from the NHANES III survey, we show how AI can clarify inter-individual variations in exposure-associated risks. The results add insights not easily obtained from traditional parametric or semi-parametric models such as logistic regression and Cox proportional hazards models, illustrating the advantages of non-parametric approaches for quantifying heterogeneous causal effects on survival times. This paper also suggests some practical implications of using AI in regulatory health risk assessments and public policy decisions.
人工智能辅助数据分析可以帮助风险分析师更好地理解暴露-反应关系，因为它能相对轻松地应用先进的统计和机器学习方法、检验其假设并解释结果。本文展示了诸如ChatGPT等大语言模型在促进健康风险评估的统计分析（包括生存数据分析）方面的潜力。通过使用相对新颖和先进的方法（如使用随机生存森林的个体条件期望（ICE）图以及使用因果生存森林估计的异质性治疗效果（HTE））进行人工智能引导的分析，可以将人群水平的暴露-反应函数分解为个体水平的暴露-反应函数。这些函数揭示了在其他变量固定的情况下，不同暴露水平个体之间风险的异质性程度。通过将这些方法应用于美国国家健康与营养检查调查（NHANES III）中从不吸烟男性的血铅水平（BLL）和死亡风险的说明性数据集，我们展示了人工智能如何阐明暴露相关风险的个体间差异。结果提供了传统参数或半参数模型（如逻辑回归和Cox比例风险模型）不易获得的见解，说明了非参数方法在量化对生存时间的异质性因果效应方面的优势。本文还提出了在监管健康风险评估和公共政策决策中使用人工智能的一些实际影响。
