Integrating Wound Images and Clinical Text for Pressure Injury Assessment and Treatment Recommendation

Area of Science:

• Artificial Intelligence in Medicine
• Computational Pathology
• Multimodal Learning

Background:

• Pressure injury management necessitates integrated decision-making for staging, debridement, and dressing selection.
• Existing computational approaches often address these tasks in isolation, using single data modalities.

Purpose of the Study:

• To develop and evaluate a multimodal, multi-task framework for pressure injury assessment.
• To integrate wound photographs and clinical text for improved decision support in wound care.

Main Methods:

• Developed a framework integrating image and text data for multi-task learning.
• Compared various models including image-only, text-only, LLM-based, fusion, and image-language models.
• Trained and validated models on public and private datasets, assessing performance across classification and recommendation tasks.

Main Results:

• The best performing model (ResNet-50 + DeepSeek-R1-Distill-Qwen-1.5B) achieved high accuracy for stage classification (82.92%) and debridement prediction (89.91%).
• The model demonstrated strong performance in recommending primary (ROUGE-1: 58.30%) and secondary (ROUGE-1: 61.06%) dressings.
• Interpretability analysis revealed clinically relevant patterns in image attention and text attribution.

Conclusions:

• Multimodal learning offers a robust approach for pressure injury assessment and treatment recommendations.
• The developed framework provides clinically aligned decision support, integrating diverse data sources for comprehensive wound care.