MacLellan Calum Robert, Petkov Hristo, McKeag Conor, Dong Feng, Lowe David John, Maguire Roma, Moschoyiannis Sotiris, Armes Jo, Skene Simon, Finlinson Alastair, Sainsbury Christopher
Department of Biomedical Engineering, University of Strathclyde, Glasgow, United Kingdom.
Department of Computer and Information Sciences, University of Strathclyde, Glasgow, United Kingdom.
PLOS Digit Health. 2025 Jul 21;4(7):e0000927. doi: 10.1371/journal.pdig.0000927. eCollection 2025 Jul.
Randomized controlled trials (RCTs) remain the benchmark for assessing treatment effects but are limited to phenotypically narrow populations by design. We introduce a novel generative artificial intelligence (AI) driven emulation method that infers effect size through virtual clinical trials, which can emulate the RCT process and potentially extrapolate into wider populations. We validate the virtual trials by comparing the predicted impact of glucagon-like peptide-1 (GLP-1) agonists on HbA1c in type-2 diabetes (T2DM) with its true efficacy established in the LEAD-5 trial. Our emulation model learns treatment effects from real-world evidence data by a combined generative AI and causal learning approach. Training data comprised pre- and post-treatment outcomes for 5,476 people with T2DM. We considered three treatment arms: GLP-1 (Liraglutide), basal insulin (glargine), and placebo. After training, virtual trials were conducted by sampling 232 virtual patients per arm (according to the LEAD-5 inclusion criteria) and predicting post-treatment outcomes. We used difference-in-differences (DiD) for pairwise comparisons between arms. Our goal was to emulate LEAD-5 by demonstrating a significant DiD in post-treatment HbA1c reduction for GLP-1 compared to basal insulin and placebo. We found significant differences in HbA1c reduction for GLP-1 vs basal insulin (-1.21 mmol/mol (-0.11%); p < 0.001) and GLP-1 vs placebo (-2.58 mmol/mol (-0.24%); p < 0.001) in our virtual populations, consistent with LEAD-5 (Liraglutide vs glargine: -2.62mmol/mol (-0.24%); p = 0.0015, Liraglutide vs placebo: -11.91 mmol/mol (-1.09%); p < 0.0001). The causal AI-powered clinical trials can emulate LEAD-5 in important measurements for T2DM. Our algorithm is specialty agnostic and can explore counterfactual questions, making it suitable for further study in the generalizability of RCT results in real-world populations to support clinical decision-making and policy recommendations.
随机对照试验（RCT）仍然是评估治疗效果的基准，但在设计上仅限于表型狭窄的人群。我们引入了一种新型的生成式人工智能（AI）驱动的模拟方法，该方法通过虚拟临床试验推断效应大小，它可以模拟RCT过程，并有可能推广到更广泛的人群。我们通过比较胰高血糖素样肽-1（GLP-1）激动剂对2型糖尿病（T2DM）患者糖化血红蛋白（HbA1c）的预测影响与其在LEAD-5试验中确定的真实疗效，来验证虚拟试验。我们的模拟模型通过结合生成式AI和因果学习方法，从真实世界证据数据中学习治疗效果。训练数据包括5476名T2DM患者的治疗前和治疗后结果。我们考虑了三个治疗组：GLP-1（利拉鲁肽）、基础胰岛素（甘精胰岛素）和安慰剂。训练后，通过每组抽取232名虚拟患者（根据LEAD-5纳入标准）并预测治疗后结果来进行虚拟试验。我们使用差异中的差异（DiD）进行组间的成对比较。我们的目标是通过证明与基础胰岛素和安慰剂相比，GLP-1治疗后HbA1c降低有显著的DiD，来模拟LEAD-5试验。我们发现在虚拟人群中，GLP-1与基础胰岛素相比，HbA1c降低有显著差异（-1.21 mmol/mol（-0.11%）；p < 0.001），GLP-1与安慰剂相比也有显著差异（-2.58 mmol/mol（-0.24%）；p < 0.001），这与LEAD-5试验结果一致（利拉鲁肽与甘精胰岛素：-2.62 mmol/mol（-0.24%）；p = 0.0015，利拉鲁肽与安慰剂：-11.91 mmol/mol（-1.09%）；p < 0.0001）。因果AI驱动的临床试验可以在T2DM的重要测量指标上模拟LEAD-5试验。我们的算法不受专业领域限制，可以探索反事实问题，使其适合进一步研究RCT结果在真实世界人群中的可推广性，以支持临床决策和政策建议。
