Machine learning models for synthesizing actionable care decisions on lower extremity wounds

Area of Science:

• Medical informatics
• Artificial intelligence in healthcare
• Wound care research

Background:

• Lower extremity chronic wounds impact millions annually, often receiving non-standardized care due to limited expert access.
• Inequitable access to wound specialists in underserved regions leads to suboptimal patient outcomes and increased healthcare costs.

Purpose of the Study:

• To evaluate machine learning (ML) classifiers for generating objective wound care decisions for chronic lower extremity wounds.
• To determine if visual wound features alone are sufficient for ML-driven decision-making.
• To assess the impact of incorporating unstructured clinical text on ML classifier accuracy.

Main Methods:

• Exploration of various ML classifiers, including single classifiers, ensemble methods (bagged & boosted), and deep learning networks.
• Training and testing classifiers on a dataset of 205 chronic wound images and associated expert clinical notes.
• Comparison of classifier performance using visual features only versus combined visual and textual data.

Main Results:

• The Gradient Boosted Machine (XGBoost) achieved the highest accuracy (81%) when utilizing both visual and textual wound features.
• A Support Vector Machine classifier achieved 76% accuracy using only visual wound features.
• Incorporating unstructured text from wound experts significantly improved decision-making accuracy.

Conclusions:

• ML classifiers can accurately generate wound care decisions for lower extremity chronic wounds, promoting standardized and objective care.
• Leveraging clinical comments from wound experts enhances the accuracy of ML-based wound care decision support systems.
• This approach represents a significant step towards improving the quality of care for chronic wound patients, especially in areas with limited specialist access.