Generative Artificial Intelligence Models inPharmacokinetics: A Study on aTwo-Compartment Population Model

Abstract This study investigates the potential of AI generative models in pharmacometrics, focusing on developing a two-compartment population pharmacokinetic (PK) model using data from Hosseini et al. (2018). Artificial Intelligence (AI) generative models assist non-professional programmers in learning programming languages, techniques, and skills relevant to PK modelling. To evaluate free AI generative models like ChatGPT 3.5 (ChatGPT), Microsoft Copilot v4.0 (Copilot), and Bard v4.0 (Gemini), PK models were constructed in R Studio, addressing essential tasks in pharmacometrics and clinical pharmacology, including model descriptions, necessary inputs, results, and code generation, focusing on reproducibility. ChatGPT and Copilot outperformed Gemini, demonstrating strong foundational knowledge, advanced concepts, and practical skills, including PK code structure and syntax. Validation showed high accuracy in estimated and simulated plots, with negligible differences in Cl and Q (6.89 ± 0.2 and 45.5 ± 17.4 [reference] vs. 7.02 ± 0.78 and 50.55 ± 0.5 [AI models]), and volumes Vc and Vp (49.15 ± 3.8 and 34.61 ± 5.2 [reference] vs. 49.86 ± 1.43 and 33.14 ± 1.33 [AI models]). Metrics results showed AFE, AAFE, and MPE values of 0.99, 1.14, and -1.85, respectively. AI generative models can accurately retrieve PK parameters from literature, develop population PK models in R, and create interactive Shiny applications for result visualization. However, recognizing the limitations and challenges of AI generative models, including accuracy, reliability, transparency, and reproducibility of AI-generated data and code, is crucial.