Objective To construct and validate a deep learning model for the automated staging of pressure inju-ries(PI).Methods A total 201 images from January 2021 to June 2023 were selected from the electronic pressure sore management system of Changshu First People's Hospital,and PI was categorized into 4 stages,including 21 im-ages of stage Ⅰ,41 images of stage Ⅱ,101 images of advanced stage,and 38 images of deep tissue injury.DenseNet121,EfficientNet,ResNet101,and ResNet50 neural networks based on the Convolutional Neural Network(CNN)framework were used to establish deep learning models for PI stratification tasks;Model evaluation indica-tors included accuracy,recall,precision,F1 score,and reading time.The reading performance of the deep learning model were compared with that of 2 nurses with different years of experience.Finally,interpretability analysis on the best-performing CNN model was conducted and pressure injury videos was performed real-time prediction.Re-sults Among 4 deep learning models in the test set,DenseNet121 demonstrated superior accuracy(0.895),followed by ResNet50(0.816),both of which were higher than the experienced nurse(0.805)and the less experienced nurse(0.756).Also,all deep learning models took less than 10 seconds to read the test set,faster than the nurses(all＞250 s).Finally,we used Gradient Weighted Class Activation Mapping(Grad-CAM)and SHAP techniques for an in-depth analysis of the optimal model,DenseNet121,highlighting the key areas in the images that significantly in-fluence the model's judgment,and achieved real-time prediction on PI videos.Conclusion A deep learning model that performs better than manual nurse evaluation was successfully established in the assessment of pressure injury risks.This computer vision-based deep learning model significantly assists nurses in conducting more accurate stratification of PI,revealing the vast potential of deep learning in clinical medicine applications.
deep learningpressure injuryartificial intelligenceconvolutional neural networks
万方数据
维普
