Automated Pressure Injury Assessment and Documentation Generation Using Vision-Language Model

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Clinical Documentation

Background:

• Pressure injury assessment and documentation are vital but time-consuming clinical tasks.
• Current inter-rater reliability for pressure injury assessment is suboptimal, ranging between 60-70%.

Purpose of the Study:

• To develop and evaluate an automated approach for pressure injury assessment and clinical description generation.
• To leverage the Florence-2 vision-language model for enhanced accuracy and efficiency in pressure injury evaluation.

Main Methods:

• Training the Florence-2 model on 946 pressure injury images, with data augmentation to 275 images per grade.
• Utilizing the model for automated grading and standardized clinical description generation of pressure injuries.

Main Results:

• Achieved robust performance across pressure injury grades with F1 scores from 73.17% to 95.00%.
• Demonstrated strong capability in generating standardized clinical descriptions, evidenced by a BERTScore of 85.58%.

Conclusions:

• The automated approach shows significant potential for improving assessment consistency and documentation efficiency in clinical practice.
• While challenges remain in differentiating Stage 3 and 4 injuries, the solution addresses critical needs for standardization and accuracy in pressure injury management.