Written by Hannah Harp
Spoon Feed
Large language models (LLMs) ChatGPT-4o and Claude-3.0 were faster and more accurate than experienced nurses when faced with different types of pediatric dose calculation scenarios.
Let’s crunch the numbers – humans vs. robots
Medication dosing errors are the most common type of medical errors, and dosing errors are even more common in inpatient pediatric medicine due to the complexities of weight-based dosing. A 2008 study showed that 22% of adverse drug events (including mild adverse reactions like nausea and pruritus) were preventable. This cross-sectional study assessed whether LLMs can enhance pediatric dosing accuracy compared to experienced nurses. Researchers tested ChatGPT-4o, Claude-3.0, and Llama 3 8B against 101 pediatric nurses using a nine-question pediatric dosing survey.
ChatGPT-4o and Claude-3.0 achieved 100% accuracy, with faster response times (15.7–75.12 s) versus nurses (mean 1621.2 s, 93.14% accuracy). Statistical analysis confirmed performance differences (p < 0.001), indicating strong potential for LLMs in reducing pediatric dosing errors. This was a relatively small study with only 101 participants. LLM accuracy may not translate to more complicated clinical settings, and by the same token, human accuracy is likely decreased in a high-pressure and chaotic clinical scenario.
How will this change my practice?
Mg/kg/day or mg/kg/dose? What is that in mL? What if you need to add diluent? Also, room 12’s line is occluded, room 8 needs help getting to the bathroom, and the other nurse is on break. No wonder dosing errors occur so frequently! Let’s take whatever help we can get.
Editor’s note: Math is hard, especially in a busy clinical setting, but while clinical uses for AI are promising, I still approach its use with caution. Once accuracy is further validated, it will be interesting to see how AI can be incorporated into work-flow or EMRs that don’t already calculate weight-based dosing. ~Kelsey Hart
Source
Can large language models assist with pediatric dosing accuracy? Pediatr Res. 2025 Mar 8. doi: 10.1038/s41390-025-03980-8. Epub ahead of print. PMID: 40057653
