Scholarlink

Background <p>Venous needle dislodgement (VND) during hemodialysis can be fatal. We hypothesized that artificial intelligence technology could be applied to detect bleeding caused by VND visually using surveillance cameras. In this study, we developed the bleeding detection model (BDM) and compared its accuracy for images with backgrounds different from those of a training set using the bleeding image classification model (BICM) to evaluate its usefulness.</p> Method <p>An experiment was conducted to reproduce the appearance of VND using simulated blood, capturing 3070 bleeding and normal images for the training set. Then, 436 bleeding images with different backgrounds and 555 normal images from 24 patients undergoing hemodialysis were collected as a test set. A BDM and a BICM were constructed using You Only Look Once version 8 and Visual Geometry Group 16, respectively. Each model was trained using the training set, and detection accuracy was evaluated using the test set.</p> Result <p>The accuracy evaluation indices of the BDM and the BICM in bleeding class detection were <i>F</i>1-score 0.92 and 0.81, precision 0.89 and 0.78, and recall 0.94 and 0.85, respectively.</p> Conclusions <p>BDM reduced background influence and improved detection accuracy. Applying an object detection model to VND monitoring using surveillance cameras is beneficial because the location of bleeding is displayed as a rectangle, facilitating the interpretation of the results.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

The usefulness of bleeding caused by needle dislodgement detection model during hemodialysis using artificial intelligence-based object detection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年6月16日

查看全文
  • Daichi Ninomiya,
  • Kohei Aoki,
  • Ryo Yoshitake,
  • Kana Teshigahara,
  • Naho Hiratani,
  • Shiori Kubota,
  • Kenji Fujii,
  • Keiko Onoue,
  • Shumpei Nakamura

摘要

Background

Venous needle dislodgement (VND) during hemodialysis can be fatal. We hypothesized that artificial intelligence technology could be applied to detect bleeding caused by VND visually using surveillance cameras. In this study, we developed the bleeding detection model (BDM) and compared its accuracy for images with backgrounds different from those of a training set using the bleeding image classification model (BICM) to evaluate its usefulness.

Method

An experiment was conducted to reproduce the appearance of VND using simulated blood, capturing 3070 bleeding and normal images for the training set. Then, 436 bleeding images with different backgrounds and 555 normal images from 24 patients undergoing hemodialysis were collected as a test set. A BDM and a BICM were constructed using You Only Look Once version 8 and Visual Geometry Group 16, respectively. Each model was trained using the training set, and detection accuracy was evaluated using the test set.

Result

The accuracy evaluation indices of the BDM and the BICM in bleeding class detection were F1-score 0.92 and 0.81, precision 0.89 and 0.78, and recall 0.94 and 0.85, respectively.

Conclusions

BDM reduced background influence and improved detection accuracy. Applying an object detection model to VND monitoring using surveillance cameras is beneficial because the location of bleeding is displayed as a rectangle, facilitating the interpretation of the results.