Intelligent Prediction of Pressure Injury by Image-Based Feature Variable With Machine Learning
Background:
Rapid and objective pressure injury (PI) assessment is crucial for preventing further wound deterioration.
Objectives:
This study aimed to develop an image-based intelligent system for PI determination that does not rely on human sensory evaluation.
Methods:
An image-based PI determination system was developed using a combination method of feature variable extraction and machine learning. Color and texture features were selected because they are closely related to human sensory evaluation methods. The digital data from these selected feature variables served as the original data set for model construction. Then, the contribution and relationships between the extracted feature variables and model performance were investigated using shapely additive explanations and Spearman algorithms to enhance the robustness of the PI determination model. Additionally, the influence of sample size and K values on model performance was determined for robust model construction.
Results:
A k -nearest neighbor algorithm was used to build PI prediction models based on these selected variables and image samples. The classification rate for the best model is 97.22% and 97.08% on the training and test sets, respectively.
Discussion:
All results demonstrate that image-based feature variables coupled with machine learning are efficient for PI determination and perhaps other medical diagnoses involving visual recognition.
