Large language models and clinical calculations: to err is human and machines are not exempt (28 May 2025)

Summary

Medication dosing errors occur frequently and contribute to preventable patient harm and negative outcomes (including numerous patient deaths each year in the US). Dosing errors are particularly common in neonatal and paediatric populations, where weight-based dosing is often required and drug formulations are commonly tailored towards adult populations. 

Hospitalised neonates require frequent dosing adjustments as their weights can change substantially over the course of their hospitalization and even day to day, increasing the potential for dosing errors. Technologies such as computerised order entry, clinical decision support systems, and electronic prescribing strategies have been used to improve dosing accuracy and prevent adverse drug events with mixed results. Additionally, paediatric functionalities are often not integrated into electronic health records (EHR) or tools tailored to the adult population are incorrectly applied to paediatric patients.

In this issue of Pediatric Research, Levin and colleagues compared the accuracy of three Large Language Models (LLMs) to nurses of varying clinical backgrounds and experience levels in calculating paediatric medication dosing.

Although this study focused on nurses, it applies to all healthcare providers. Medication dosing errors do not occur in a vacuum and it is the responsibility of all healthcare providers (including nurses, physicians, pharmacists, technicians, etc.) to ensure that medications are given at the correct dose, route, interval, and duration.