Real-time hand rubbing quality estimation using deep learning enhanced by separation index and feature-based...

Area of Science:

• Computer Science
• Healthcare Technology
• Artificial Intelligence

Background:

• Hand hygiene is critical for preventing healthcare-associated infections and controlling disease transmission, including COVID-19.
• The World Health Organization (WHO) provides a 12-step guideline for effective alcohol-based hand rubbing.
• Ensuring consistent compliance with hand hygiene protocols is challenging in clinical practice.

Purpose of the Study:

• To develop an automated system for evaluating hand rubbing technique quality.
• To build a reliable hand hygiene system using deep learning.
• To introduce a novel metric for assessing model confidence in predictions.

Main Methods:

• Collected a large-scale dataset of real and fake hand rubbing motions.
• Utilized a swift Separation Index (SI) method to compare pre-trained deep learning networks without fine-tuning.
• Developed a Feature-Based Confidence (FBC) metric to differentiate model performance.

Main Results:

• The Inception-ResNet architecture demonstrated the highest SI, achieving 98% accuracy after fine-tuning.
• A lightweight version of the Inception-ResNet model was developed with comparable accuracy.
• The Inception-ResNet model showed higher confidence (5%) than its lightweight counterpart, despite being slower.

Conclusions:

• A deep learning-based method effectively qualifies the hand rubbing process, addressing real-time application needs.
• The developed system, DeepHARTS, can be deployed in various organizations and healthcare environments to monitor hand hygiene quality.
• The Feature-Based Confidence metric provides a valuable tool for selecting superior models in similar accuracy scenarios.