Researchers in Sweden’s Region Västmanland have developed an AI system capable of identifying medication errors in hospital incident reports with 96% accuracy, matching the performance of expert pharmacists. The study, published in JAMIA Open, suggests that Large Language Models (LLMs) can solve the chronic problem of inconsistent and under-reported safety data in healthcare.
According to the report, medication errors cause roughly 10% of preventable patient harm, but traditional manual reporting systems often capture only a fraction of these incidents due to inconsistent human classification. To solve this problem, the research team used OpenAI’s O4-mini model to analyze 75,000 anonymized reports. The AI’s high performance was achieved by “encoding tacit knowledge” which means incorporating the informal expertise and professional intuition of veteran pharmacists into the model’s instructions.
The researchers used a specific architecture called “Critic Model Integration”, where a secondary AI reviews the initial classification to ensure reliability. If the AI detects an ambiguity, the case is flagged for human review.
By automating the sorting of thousands of unstructured reports, this technology allows hospitals to identify systemic risks in real-time, moving more quickly toward targeted strategies that prevent patient harm and reduce the billions in associated healthcare costs.
