OBJECTIVE: To investigate the accuracy and reliability of AI chatbots to estimate pharmacokinetic parameters from limited patient samples and population data for potential application in teaching Bayesian concepts. METHODS: Two plasma concentration-time data after a single intravenous dose, along with population values for volume of distribution (V) and elimination rate constant (k), were entered into free versions of ChatGPT and Gemini. Three prompts were engineered to assess and improve the accuracy and consistency of patient-only (based on plasma concentrations) and Bayesian (based on both plasma concentrations and population data) estimates of V and k. To assess consistency, each prompt was analyzed seven times across three separate occasions, resulting in 21 replicates per chatbot. The optimized prompt was then used to create a simulation assignment comprising three scenarios with varying levels of assay and/or population parameter variability. RESULTS: Initially, both chatbots produced inconsistent and sometimes inaccurate responses for patient-only and Bayesian estimates of V and k. Prompt engineering improved estimates of pharmacokinetic parameters for both chatbots. However, ChatGPT produced more accurate and consistent results for both patient-only and Bayesian estimates with the optimized prompt. The simulation assignment with the optimized prompt revealed that the simulation scenarios accurately and reliably predict the effects of changing the assay and population parameters variability on the Bayesian estimates of V and k. CONCLUSION: Free versions of chatbots can serve as alternatives to specialized pharmacokinetic software, making advanced pharmacokinetic simulation tools, including those for Bayesian analysis, more accessible to pharmacy educators and students.
Reza Mehvar (Fri,) studied this question.