Assessing the accuracy of 4 generative AI tools to guide dosing of long-acting injectable antipsychotics

The purpose of this study was to review the accuracy of 4 different artificial intelligence (AI) tools in providing dosing recommendations for initiation and missed doses of long-acting injectable antipsychotics (LAIAs).

The increasing number of LAIAs provides a valuable treatment option for individuals with schizophrenia and bipolar disorder who struggle with adherence. Currently, there are over 10 second-generation LAIA formulations available in the United States; however, dosing and administration can be complex, leading to the possibility of errors. As the use of AI continues to expand in healthcare, clinicians may utilize AI for guidance in dosing LAIAs. Our study evaluated the accuracy of 4 generative AI tools (ChatGPT, Microsoft Copilot Chat, Perplexity, and OpenEvidence) in providing dosing recommendations for both initiation and missed-dosing scenarios for LAIAs. Accuracy rates among the models varied significantly. OpenEvidence was the most accurate, providing an inacurate response to only one query (a 95% accuracy rate). ChatGPT was the least accurate, with a 40% accuracy rate. Most of the inaccuracies (62%) occurred when the AI tools provided recommendations related to missed-dose scenarios.

This study highlights the importance of providers using caution when relying on AI tools to provide dosing recommendations for patients prescribed LAIAs. Tools that are rooted in more scientific data, like OpenEvidence, may be more useful in providing accurate information quickly than conversation-based tools, like ChatGPT. If utilizing AI tools for guidance on medication dosing and recommendations, providers should be sure to also review recommendations from package inserts or other reliable drug information sources.