Modeling anesthetic-drug label detection for low-resource operating rooms

Abstract

Medication errors during surgical procedures pose significant risks to patient safety, particularly in low-resource healthcare settings. In this study, we present the development and evaluation of computer vision models for detecting clinical errors related to syringe and vial swaps during anesthetic drug administration at the Comprehensive Rehabilitation Services for People with Disability in Uganda (CoRSU Hospital). Leveraging data collected through wearable head-mounted cameras, we created a custom dataset comprising annotated images of syringes and vials labeled with a diverse range of drugs used in the operating room. Three deep learning-based object detection architectures were trained: YOLOv11n, YOLOv12n, and RT-DETR large. The RT-DETR model achieved marginally higher performance, with a precision of 0.89, recall of 0.85, and mAP50 of 0.86, outperforming both YOLO variants across key metrics. The models provide a foundation for real-time monitoring tools aimed at reducing medication errors and improving clinical safety protocols in resource-constrained surgical environments.