Photographic LVAD Driveline Wound Infection Recognition Using Deep Learning

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Medical Imaging Analysis

Background:

• Increasing use of mechanical heart support devices (LVAD) necessitates robust remote patient monitoring solutions.
• Driveline exit site infections are a common complication in LVAD patients, requiring timely and accurate assessment.
• Current remote monitoring methods may lack objective data for precise infection severity evaluation.

Purpose of the Study:

• To investigate machine learning (ML) approaches for automated infection severity recognition from driveline exit site images.
• To evaluate the performance of a U-net convolutional neural network (CNN) for image segmentation and a classification CNN for infection grading.
• To explore the potential of incorporating additional imaging modalities to enhance diagnostic accuracy.

Main Methods:

• Driveline tube segmentation was performed using a U-net CNN, achieving a Dice score of 0.95.
• A separate classification CNN was trained to categorize driveline exit site images into three infection severity classes.
• Model performance was evaluated against expert assessments of the same image set.

Main Results:

• The segmentation U-net CNN demonstrated high performance with a Dice score of 0.95.
• The infection classification CNN achieved an overall accuracy of 67%, comparable to human expert performance.
• The study suggests that visual information alone may be insufficient for differentiating infection severities accurately.

Conclusions:

• Machine learning, particularly CNNs, shows promise for remote monitoring of LVAD driveline infections.
• Current image-based ML models achieve performance levels similar to human experts, indicating limitations in visual data.
• Integrating complementary data, such as thermographic imaging, is recommended to improve the accuracy of infection severity assessment in remote patient monitoring.