# A labeled dataset of simulated phlebotomy procedures for medical AI: polygon annotations for object detection and human-object interaction

**Authors:** Raúl Jiménez Cruz, César Torres-Huitzil, Marco Franceschetti, Ronny Seiger, Luciano García-Bañuelos, Barbara Weber  
**arXiv:** [2602.04624](https://arxiv.org/abs/2602.04624)  
**Date:** 2026-02-04  
**Summary page:** https://www.summarizepaper.com/en/paper/2602.04624v1/

## Key points

- Dataset contains 11,884 labeled images of a simulated blood extraction procedure on a training arm
- Images were selected from high-definition videos and curated to eliminate redundancy using SSIM filtering
- An automated face-anonymization process was implemented prior to frame selection
- Each image is annotated with polygon outlines for five medical classes: syringe, rubber band, disinfectant wipe, gloves, and training arm
- Annotations are compatible with object detection frameworks like YOLOv8
- Dataset is divided into training (70%), validation (15%), and test (15%) subsets
- Primary objective is to advance research in medical training automation and human-object interaction
- Tasks enabled by the dataset include phlebotomy tool detection, procedural step recognition, workflow analysis, conformance checking, and creation of educational systems with structured feedback
- Data and labels are available on Zenodo for transparency and accessibility to researchers

## Summary

This data article presents a comprehensive dataset of 11,884 labeled images capturing a simulated blood extraction (phlebotomy) procedure conducted on a training arm. The images were meticulously selected from high-definition videos recorded in controlled settings and underwent curation to eliminate redundancy using Structural Similarity Index Measure (SSIM) filtering. Prior to frame selection, an automated face-anonymization process was implemented across all videos. Each image in the dataset is annotated with polygon outlines for five essential medical classes: syringe, rubber band, disinfectant wipe, gloves, and the training arm itself. These annotations were exported in a segmentation format compatible with modern object detection frameworks like YOLOv8, ensuring broad applicability for various research purposes. The dataset is divided into three subsets: training (70%), validation (15%), and test (15%), facilitating robust model development and evaluation. Its primary objective is to advance research in medical training automation and human-object interaction by enabling tasks such as phlebotomy tool detection, procedural step recognition, workflow analysis, conformance checking, and the creation of educational systems that offer structured feedback to medical trainees. The availability of both the data and accompanying label files on Zenodo promotes transparency and accessibility for researchers interested in leveraging this resource. Authored by Raúl Jiménez Cruz, César Torres-Huitzil, Marco Franceschetti,
Ronny Seiger,Luciano García-Bañuelos,and Barbara Weber,this dataset titled "A labeled dataset of simulated phlebotomy procedures for medical AI: polygon annotations for object detection and human-object interaction" serves as a valuable asset for advancing AI-driven solutions in healthcare education and practice.